Privacy in Machine Learning (NeurIPS 2021 Workshop)

The past decade has seen data privacy emerge as a fundamental and pressing issue. Among the tools developed to tackle it, differential privacy has emerged as a central and principled framework, with specific variants capturing various threat models. In particular, the recently proposed shuffle model of differential privacy allows for promising tradeoffs between accuracy and privacy. However, the shuffle model may not be suitable in all situations, as it relies on a distributed setting where all users can coordinate and trust (or simulate) a joint shuffling algorithm. To address this, we introduce a new model, the group shuffle model, in which users are partitioned into several groups, each group having its own local shuffler. We investigate the privacy/accuracy tradeoffs in our model, by comparing it to both the shuffle and local models of privacy, which it some sense interpolates between. In addition to general relations between group shuffle, shuffle, and local privacy, we provide a detailed comparison of the cost and benefit of the group shuffle model, by providing both upper and lower bounds for the specific task of binary summation.

Clustering is a popular unsupervised machine learning technique that groups similar input elements into clusters. In many applications, sensitive information is clustered that should not be leaked. Moreover, nowadays it is often required to combine data from multiple sources to increase the quality of the analysis as well as to outsource complex computation to powerful cloud servers. This calls for efficient privacy-preserving clustering. In this work, we systematically analyze the state-of-the-art in privacy-preserving clustering. We implement and benchmark today's four most efficient fully private clustering protocols by Cheon et al. (SAC'19), Meng et al. (ArXiv'19), Mohassel et al. (PETS'20), and Bozdemir et al. (ASIACCS'21) with respect to communication, computation, and clustering quality.

Differential privacy (DP) offers strong protection against adversaries with arbitrary side-information and computational power. However, many implementations of DP mechanisms leave themselves vulnerable to side channel attacks, such as timing attacks. As many privacy mechanisms, such as the exponential mechanism, do not lend themselves to easy implementations, when sampling methods such as MCMC or rejection sampling are used, the runtime can leak privacy. In this work, we quantify the privacy cost due to the runtime of a rejection sampler in terms of -DP. We also propose three modifications to the rejection sampling algorithm, to protect against timing attacks by making the runtime independent of the data. We also use our techniques to develop an adaptive rejection sampler for log-Holder densities, which also has data-independent runtime.

Accepted Papers

Antti Koskela, Mikko A. Heikkilä, Antti Honkela
Tight Accounting in the Shuffle Model of Differential Privacy [openreview] [Visit Poster at Spot A2]

Shuffle model of differential privacy is a novel distributed privacy model based on a combination of local privacy mechanisms and a trusted shuffler. It has been shown that the additional randomisation provided by the shuffler improves privacy bounds compared to the purely local mechanisms. Accounting tight bounds, especially for multi-message protocols, is complicated by the complexity brought by the shuffler. The recently proposed Fourier Accountant for evaluating -differential privacy guarantees has been shown to give tighter bounds than commonly used methods for non-adaptive compositions of various complex mechanisms. In this paper we show how to compute tight privacy bounds using the Fourier Accountant for multi-message versions of several ubiquitous mechanisms in the shuffle model and demonstrate looseness of the existing bounds in the literature.

Akash Bharadwaj, Graham Cormode
Sample-and-threshold differential privacy: Histograms and applications [openreview] [Visit Poster at Spot A3]

Federated analytics aims to compute accurate statistics from distributed datasets. A "Differential Privacy" (DP) guarantee is usually desired by the users of the devices storing the data. In this work, we prove a strong (ε, δ)-DP guarantee for a highly practical sampling-based procedure to derive histograms. We also provide accuracy guarantees and show how to apply the procedure to estimate quantiles and modes.

Cangxiong Chen, Neill D. F. Campbell
Understanding Training-Data Leakage from Gradients in Neural Networks for ImageClassifications [openreview] [Visit Poster at Spot B0]

Federated learning of deep learning models for supervised tasks, e.g. image classification and segmentation, has found many applications: for example in human-in-the-loop tasks such as film post-production where it enables sharing of domain expertise of human artists in an efficient and effective fashion. In many such applications, we need to protect the training data from being leaked when gradients are shared in the training process due to IP or privacy concerns. Recent works have demonstrated that it is possible to reconstruct the training data from gradients for an image-classification model when its architecture is known. However, there is still an incomplete theoretical understanding of the efficacy and failure of such attacks. In this paper, we analyse the source of training-data leakage from gradients. We formulate the problem of training data reconstruction as solving an optimisation problem iteratively for each layer. The layer-wise objective function is primarily defined by weights and gradients from the current layer as well as the output from the reconstruction of the subsequent layer, but it might also involve a ‘pull-back’ constraint from the preceding layer. Training data can be reconstructed when we solve the problem backward from the output of the network through each layer. Based on this formulation, we are able to attribute the potential leakage of the training data in a deep network to its architecture. We also propose a metric to measure the level of security of a deep learning model against gradient-based attacks on the training data.

Abhinav Aggarwal, Shiva Kasiviswanathan, Zekun Xu, Oluwaseyi Feyisetan, Nathanael Teissier
Reconstructing Test Labels from Noisy Loss Scores (Extended Abstract) [openreview] [Visit Poster at Spot B1]

Label inference was recently introduced as the problem of reconstructing the ground truth labels of a private dataset from just the (possibly perturbed) cross-entropy loss scores evaluated at carefully crafted prediction vectors. In this paper, we generalize this result to provide necessary and sufficient conditions under which label inference is possible from a broad class of loss functions. We show that for many commonly used loss functions, including linearly decomposable losses, some Bregman divergence-based losses and when common activation functions are used, it is possible to design such attacks for arbitrary noise levels. We demonstrate that these attacks can also be carried out through a lightweight augmentation to any neural network model, enabling the adversary to make these attacks look benign. Our results call to attention these vulnerabilities which might be currently under silent exploitation. Armed with this information, individuals and organizations, which vend these seemingly innocuous aggregate metrics from their classification models, can grasp the potential scope of the resulting information leakage.

Ahmed Roushdy Elkordy, Saurav Prakash, Salman Avestimehr
Basil: A Fast and Byzantine-Resilient Approach for Decentralized Training [openreview] [Visit Poster at Spot C0]

Decentralized (i.e., serverless) learning across a large number of distributed nodes (e.g., mobile users) has seen a surge of recent interests. The key advantage of these setups is that they provide privacy for the local data of the users while not requiring a server for coordinating the training. They can, however, suffer substantially from potential Byzantine nodes in the network who can degrade the training performance. Detection and mitigation of Byzantine behaviors in a decentralized learning setting is a daunting task, especially when the data distribution at the users is heterogeneous. As our main contribution, we propose \texttt{Basil}, a fast and computationally efficient Byzantine robust algorithm for decentralized training systems, which leverages a novel sequential, memory assisted and performance based criteria for training over a logical ring while filtering the Byzantine users. In the IID dataset distribution setting, we provide the theoretical convergence guarantees of \texttt{Basil}, demonstrating its linear convergence rate. Furthermore, for the IID setting, we experimentally demonstrate that \texttt{Basil} is robust to various Byzantine attacks, including the strong Hidden attack, while providing up to 16% higher test accuracy over the state-of-the-art Byzantine-resilient decentralized learning approach. Additionally, we generalize \texttt{Basil} to the non-IID dataset distribution setting by proposing Anonymous Cyclic Data Sharing (ACDS), a technique that allows each node to anonymously share a random fraction of its local non-sensitive dataset (e.g., landmarks images) with all other nodes. We demonstrate that \texttt{Basil} alongside ACDS with only 5% data sharing provides effective toleration of Byzantine nodes, unlike the state-of-the-art Byzantine robust algorithm that completely fails in the heterogeneous data setting.

César Sabater, Jan Ramon
Zero Knowledge Arguments for Verifiable Sampling [openreview] [Visit Poster at Spot C1]

In privacy-preserving machine learning, it is less obvious to verify correct behavior of participants because they are not supposed to reveal their inputs in cleartext to other participants. It is hence important to make federated machine learning robust against data poisoning and related attacks. While input data can be related to a distributed ledger (blockchain), a less studied input is formed by the random sampling parties perform. In this paper, we describe strategies based on zero knowledge proofs to allow parties to prove they perform sampling (and other computations) correctly. We sketch a number of alternative ways to implement our idea and provide some preliminary experimental results.

Ossi Räisä, Antti Koskela, Antti Honkela
Differentially Private Hamiltonian Monte Carlo [openreview] [Visit Poster at Spot C2]

We present DP-HMC, a variant of Hamiltonian Monte Carlo (HMC) that is differentially private (DP). We use the penalty algorithm of Yildirim and Ermis to make the acceptance test private, and add Gaussian noise to the gradients of the target distribution to make the HMC proposal private. Our main contribution is showing that DP-HMC has the correct invariant distribution, and is ergodic. We also compare DP-HMC with the existing penalty algorithm, as well as DP-SGLD and DP-SGNHT.

Alexandra Peste, Dan Alistarh, Christoph H Lampert
SSSE: Efficiently Erasing Samples from Trained Machine Learning Models [openreview] [Visit Poster at Spot C3]

The availability of large amounts of user-provided data has been key to the success of machine learning for many real-world tasks. Recently, an increasing awareness has emerged that users should be given more control about how their data is used. In particular, users should have the right to prohibit the use of their data for training machine learning systems, and to have it erased from already trained systems. While several sample erasure methods have been proposed, all of them have drawbacks which have prevented them from gaining widespread adoption. In this paper, we propose an efficient and effective algorithm, SSSE, for samples erasure that is applicable to a wide class of machine learning models. From a second-order analysis of the model's loss landscape we derive a closed-form update step of the model parameters that only requires access to the data to be erased, not to the original training set. Experiments on CelebFaces attributes (CelebA) and CIFAR10, show that in certain cases SSSE can erase samples almost as well as the optimal, yet impractical, gold standard of training a new model from scratch with only the permitted data.

Borja Balle, Giovanni Cherubin, Jamie Hayes
Reconstructing Training Data with Informed Adversaries [openreview] [Visit Poster at Spot D0]

Given access to a machine learning model, can an adversary reconstruct the model’s training data? This work proposes a formal threat model to study this question, shows that reconstruction attacks are feasible in theory and in practice, and presents preliminary results assessing how different factors of standard machine learning pipelines affect the success of reconstruction. Finally, we empirically evaluate what levels of differential privacy suffice to prevent reconstruction attacks.

Felix Morsbach, Tobias Dehling, Ali Sunyaev
Architecture Matters: Investigating the Influence of Differential Privacy on Neural Network Design [openreview] [Visit Poster at Spot D1]

We explore the relationship between neural network architectures and model accuracy under differential privacy constraints. Our findings show that architectures that perform well without differential privacy, do not necessarily do so with differential privacy. This shows that extant knowledge on neural network architecture design cannot be seamlessly translated into the differential privacy context. Moreover, as neural architecture search consumes privacy budget, future research is required to better understand the relationship between neural network architectures and model accuracy to enable better architecture design choices under differential privacy constraints.

Oualid Zari, Chuan Xu, Giovanni Neglia
Efficient passive membership inference attack in federated learning [openreview] [Visit Poster at Spot E3]

In cross-device federated learning (FL) setting, clients such as mobiles cooperate with the server to train a global machine learning model, while maintaining their data locally. However, recent work shows that client's private information can still be disclosed to an adversary who just eavesdrops the messages exchanged between the client and the server. For example, the adversary can infer whether the client owns a specific data instance, which is called a passive membership inference attack. In this paper, we propose a new passive inference attack that requires much less computation power and memory than existing methods. Our empirical results show that our attack achieves a higher accuracy on CIFAR100 dataset (mora than 4 percentage points) with three orders of magnitude less memory space and five orders of magnitude less calculations.

Borja Rodríguez Gálvez, Filip Granqvist, Rogier van Dalen, Matt Seigel
Enforcing fairness in private federated learning via the modified method of differential multipliers [openreview] [Visit Poster at Spot F0]

Federated learning with differential privacy, or private federated learning, provides a strategy to train machine learning models while respecting users’ privacy. However, differential privacy can disproportionately degrade the performance of the models on under-represented groups, as these parts of the distribution are difficult to learn in the presence of noise. Existing approaches for enforcing fairness to machine learning models have considered the centralized setting, in which the algorithm has access to the users’ data. This paper introduces an algorithm to enforce group fairness in private federated learning, where users’ data does not leave their devices. First, the paper extends the modified method of differential multipliers to empirical risk minimization with fairness constraints, thus providing an algorithm to enforce fairness in the central setting. Then, this algorithm is extended to the private federated learning setting. The proposed algorithm, FPFL, is tested on a federated version of the Adult dataset and an “unfair” version of the FEMNIST dataset. The experiments on these datasets show how private federated learning accentuates unfairness in the trained models, and how FPFL is able to mitigate such unfairness.

Xinyu Tang, Saeed Mahloujifar, Liwei Song, Virat Shejwalkar, Milad Nasr, Amir Houmansadr, Prateek Mittal
A Novel Self-Distillation Architecture to Defeat Membership Inference Attacks [openreview] [Visit Poster at Spot F3]

Membership inference attacks are a key measure to evaluate privacy leakage in machine learning (ML) models, which aim to distinguish training members from non-members by exploiting differential behavior of the models on member and non-member inputs. We propose a new framework to train privacy-preserving models that induces similar behavior on member and non-member inputs to mitigate practical membership inference attacks. Our framework, called SELENA, has two major components. The first component and the core of our defense, called Split-AI, is a novel ensemble architecture for training. We prove that our Split-AI architecture defends against a large family of membership inference attacks, however, it is susceptible to new adaptive attacks. Therefore, we use a second component in our framework called Self-Distillation to protect against such stronger attacks, which (self-)distills the training dataset through our Split-AI ensemble and has no reliance on external public datasets. We perform extensive experiments on major benchmark datasets and the results show that our approach achieves a better trade-off between membership privacy and utility compared to previous defenses.

YUEFENG PENG
Unsupervised Membership Inference Attacks Against Machine Learning Models [openreview] [Visit Poster at Spot H0]

As a form of privacy leakage for machine learning (ML), membership inference (MI) attacks aim to infer whether given data samples have been used to train a target ML model. Existing state-of-the-art MI attacks in black-box settings adopt a so-called shadow model to perform transfer attacks. Such attacks achieve high inference accuracy but have many adversarial assumptions, such as having a dataset from the same distribution as the target model’s training data and knowledge of the target model structure. We propose a novel MI attack, called UMIA, which probes the target model in an unsupervised way without any shadow model. We relax all the adversarial assumptions above, demonstrating that MI attacks are applicable without any knowledge about the target model and its training set. We empirically show that, with far fewer adversarial assumptions and computational resources, UMIA can perform on bar with the state-of-the-art supervised MI attack.

Terrance Liu, Giuseppe Vietri, Steven Wu
Iterative Methods for Private Synthetic Data: Unifying Framework and New Methods [openreview] [Visit Poster at Spot G0]

We study private synthetic data generation for query release, where the goal is to construct a sanitized version of a sensitive dataset, subject to differential privacy, that approximately preserves the answers to a large collection of statistical queries. We first present an algorithmic framework that unifies a long line of iterative algorithms in the literature. Under this framework, we propose two new methods. The first method, private entropy projection (PEP), can be viewed as an advanced variant of MWEM that adaptively reuses past query measurements to boost accuracy. Our second method, generative networks with the exponential mechanism (GEM), circumvents computational bottlenecks in algorithms such as MWEM and PEP by optimizing over generative models parameterized by neural networks, which capture a rich family of distributions while enabling fast gradient-based optimization. We demonstrate that PEP and GEM empirically outperform existing algorithms. Furthermore, we show that GEM nicely incorporates prior information from public data while overcoming limitations of PMW^Pub, the existing state-of-the-art method that also leverages public data.

Cecilia Ferrando, Jennifer Gillenwater, Alex Kulesza
Combining Public and Private Data [openreview] [Visit Poster at Spot G1]

Differential privacy is widely adopted to provide provable privacy guarantees in data analysis. We consider the problem of combining public and private data (and, more generally, data with heterogeneous privacy needs) for estimating aggregate statistics. We introduce a mixed estimator of the mean optimized to minimize the variance. We argue that our mechanism is preferable to techniques that preserve the privacy of individuals by subsampling data proportionally to the privacy needs of users. Similarly, we present a mixed median estimator based on the exponential mechanism. We compare our mechanisms to the methods proposed in Jorgensen et al. [2015]. Our experiments provide empirical evidence that our mechanisms often outperform the baseline methods.

Arpita Patra, Thomas Schneider, Ajith Suresh, Hossein Yalame
ABY2.0: New Efficient Primitives for STPC with Applications to Privacy in Machine Learning (Extended Abstract) [openreview] [Visit Poster at Spot G2]

In this work, we improve semi-honest secure two-party computation (STPC) over rings, specially for privacy-preserving machine learning, with a focus on the efficiency of the online phase. We construct efficient protocols for several privacy-preserving machine learning (PPML) primitives such as scalar product, matrix multiplication, ReLU, and maxpool. The online communication of our scalar product is two ring elements {\em irrespective} of the vector dimension, which is a feature achieved for the first time in PPML literature. We implement and benchmark training and inference of Logistic Regression and Neural Networks over LAN and WAN networks. For training, we improve online runtime (both for LAN and WAN) over SecureML (Mohassel et al., IEEE S\&P'17) in the range 1.5×--6.1×, while for inference, the improvements are in the range of 2.5×--754.3×.

Virat Shejwalkar, Huseyin A Inan, Amir Houmansadr, Robert Sim
Membership Inference Attacks Against NLP Classification Models [openreview] [Visit Poster at Spot H0]

The success of natural language processing (NLP) is making NLP applications commonplace. Unfortunately, recent research has shown that privacy might be at stake given that these models are often trained on private user data. While privacy risks are demonstrated in text generation settings, privacy risks of the text classification settings, which subsume myriad downstream applications, are largely unexplored. In this work, we study the susceptibility of NLP classification models, used for text classification tasks, to membership inference (MI), which is a fundamental type of privacy leakage. We design a comprehensive suite of attacks to assess the risk of sample-level MI, as well as that of relatively unexplored user-level MI. We introduce novel user-level MI attacks that outperform the existing attacks and conduct experiments on Transformer-based and RNN-based NLP models. Our evaluations show that user-level MI is significantly stronger than sample-level MI. We further perform in-depth analyses showing the effect of various NLP-specific parameters on MI against NLP classification models.

Yongqin Wang, Edward Suh, Wenjie Xiong, Brian Knott, Benjamin Lefaudeux, Murali Annavaram, Hsien-Hsin Lee
Characterizing and Improving MPC-based Private Inference for Transformer-based Models [openreview] [Visit Poster at Spot H2]

Secure multi-party computation (MPC) is gaining popularity with the growing demand for privacy-preserving cloud services. While there has been plenty of attention to MPCs for convolution neural networks (CNNs), MPC-based private inference for Transformer models has not been studied in detail. This paper provides a characterization study of the performance overhead for running Transformer models with secure MPC, and proposes an optimization for embedding tables. Our study shows that Transformers introduce a couple of new challenges for MPC-based private inference: softmax and embedded tables. To address the overhead of embedding table accesses under MPC, we propose to use tensor-train (TT) decomposition, a mechanism that splits a large embedding tables into multiple smaller embedding tables. For the NLP workloads, the experiments show that the TT decomposition can speed up embedding table accesses by 2x with only a 1.19 drop in the masked-language model perplexity score.

Shubham Jain, Ana-Maria Cretu, Yves-Alexandre de Montjoye
Adversarial Detection Avoidance Attacks: Evaluating the robustness of perceptual hashing-based client-side scanning [openreview] [Visit Poster at Spot H3]

End-to-end encryption (E2EE) in messaging platforms enables people to securely and privately communicate with one another. Its widespread adoption however raised concerns that illegal content might now be shared undetected. Client-side scanning based on perceptual hashing has been recently proposed by governments and researchers to detect illegal content in E2EE communications. We propose the first framework to evaluate the robustness of perceptual hashing-based client-side scanning to detection avoidance attacks and show current systems to not be robust. We propose three adversarial attacks---a general black-box attack and two white-box attacks for discrete cosine transform-based algorithms--against perceptual hashing algorithms. In a large-scale evaluation, we show perceptual hashing-based client-side scanning mechanisms to be highly vulnerable to detection avoidance attacks in a black-box setting, with more than 99.9\% of images successfully attacked while preserving the content of the image. We further show several mitigation strategies, such as expanding the database with hashes of images modified using our attack, or increasing the detection threshold, to be ineffective against our attack. Taken together, our results shed serious doubts on the robustness of perceptual hashing-based client-side scanning mechanisms currently proposed by governments, organizations, and researchers around the world.

Andres Munoz medina, Matthew Joseph, Jennifer Gillenwater, Mónica Ribero
A Joint Exponential Mechanism for Differentially Private Top-k Set [openreview] [Visit Poster at Spot I0]

We present a novel differentially private algorithm for releasing the set of k elements with the highest counts from a data domain of d elements. We define a ``joint'' instance of the exponential mechanism (EM) whose output space consists of all O(d^k) size-k subsets; yet, we are able to show how to sample from this EM in only time O(dk^3). Experiments suggest that this joint approach can yield utility improvements over the existing state of the art for small problem sizes.

Karan Chadha, John Duchi, Rohith Kuditipudi
Private Confidence Sets [openreview] [Visit Poster at Spot I1]

We consider statistical inference under privacy constraints. In particular, we give differentially private algorithms for estimating coverage probabilities and computing valid confidence sets, and prove upper bounds on the error of our estimates and the length of our confidence sets. Our bounds apply to broad classes of data distributions and statistics of interest, and for fixed we match the higher-order asymptotic accuracy of the standard (non-private) non-parametric bootstrap.

Sen Yuan, Milan Shen, Ilya Mironov, Anderson Nascimento
Label Private Deep Learning Training based on Secure Multiparty Computation and Differential Privacy [openreview] [Visit Poster at Spot I2]

Secure Multiparty Computation (MPC) is an invaluable tool for training machine learning models when the training data cannot be directly accessed by the model trainer. Unfortunately, complex algorithms, such as deep learning models, have their computational complexities increased by orders of magnitude when performed using MPC protocols. In this contribution, we study how to efficiently train an important class of machine learning problems by using MPC where features are known by one of the computing parties and only the labels are private. We propose new protocols combining differential privacy (DP) and MPC in order to privately and efficiently train a deep learning model in such scenario. More specifically, we release differentially private information during the MPC computation to dramatically reduce the training time. All released information idoes not compromise the privacy of the labels at the individual level. Our protocols can have running times that are orders of magnitude better than a straightforward use of MPC at a moderate cost in model accuracy.

Margarita Vinaroz, Mijung Park, Mijung Park
DP-SEP: Differentially private stochastic expectation propagation [openreview] [Visit Poster at Spot I3]

We are interested in privatizing an approximate posterior inference algorithm called Expectation Propagation (EP). EP approximates the posterior by iteratively refining approximations to the local likelihoods, and is known to provide better posterior uncertainties than those by variational inference. However, using EP for large-scale datasets imposes a challenge in terms of memory requirements as it needs to maintain each of the local approximates in memory. To overcome this problem, stochastic expectation propagation (SEP) was proposed, which only considers a unique local factor that captures the average effect of each likelihood term to the posterior and refines it in a way analogous to EP. Therefore in this work, we focus on developing a differentially private stochastic expectation propagation(DP-SEP) algorithm, which outputs differentially private natural parameters of the exponential-family posteriors in each step of SEP.

Rachel Cummings, Vitaly Feldman, Audra McMillan, Kunal Talwar
Mean Estimation with User-level Privacy under Data Heterogeneity [openreview] [Visit Poster at Spot J0]

A key challenge for data analysis in the federated setting is that user data is heterogeneous, i.e., it cannot be assumed to be sampled from the same distribution. Further, in practice, different users may possess vastly different number of samples. In this work we propose a simple model of heterogeneous user data that differs in both distribution and quantity of data, and we provide a method for estimating the population-level mean while preserving user-level differential privacy. We demonstrate near asymptotic optimality of our estimator among nearly unbiased estimators. In particular, while the optimal non-private estimator can be shown to be linear, we show that privacy constrains us to use a non-linear estimator.

Nitin Agrawal, James Bell, Adrià Gascón, Matt Kusner
Certified Predictions using MPC-Friendly Publicly Verifiable Covertly Secure Commitment [openreview] [Visit Poster at Spot J1]

We address the problem of efficiently and securely enabling certified predictions on deep learning . This addresses the scenario where a party P1 owns a confidential model that has been certified by an authority to have a certain property e.g. fairness. Subsequently, another party P2 wants to perform a prediction on the model with an assurance that the certified model was used. We present a solution for this problem based on MPC commitments. Our constructions operate in the publicly verifiable covert (PVC) security model, which is a relaxation of the malicious model of MPC, appropriate in settings where P1 faces a reputational harm if caught cheating. We introduce the notion of a PVC commitment scheme and indexed hash functions to build commitment schemes tailored to the PVC framework, and propose constructions for both arithmetic and Boolean circuits that result in very efficient circuits. From a practical standpoint, our constructions for Boolean circuits are 60x faster to evaluate securely, and use 36x less communication than baseline methods based on hashing. Moreover, we show that our constructions are tight in terms of required non-linear operations, and present a technique to amplify the security properties of our constructions that allows to efficiently recover malicious guarantees with statistical security.

Edward Vendrow, Joshua Vendrow
Realistic Face Reconstruction from Deep Embeddings [openreview] [Visit Poster at Spot J2]

Modern face recognition systems use deep convolution neural networks to extract latent embeddings from face images. Since basic arithmetic operations on embeddings are needed to make comparisons, generic encryption schemes cannot be used. This leaves facial embedding unprotected and susceptible to privacy attacks that reconstruction facial identity. We propose a search algorithm on the latent vector space of StyleGAN to find a matching face. Our process yields latent vectors that generate face images that are high-resolution, realistic, and reconstruct relevant attributes of the original face. Further, we demonstrate that our process is capable of fooling FaceNet, a state-of-the-art face recognition system.

Wei-Ning Chen, Christopher A. Choquette-Choo, Peter Kairouz
Communication Efficient Federated Learning with Secure Aggregation and Differential Privacy [openreview] [Visit Poster at Spot J3]

Optimizing the \puc tradeoff is a key challenge for federated learning. Under distributed differential privacy (DP) via secure aggregation (SecAgg), we prove that the worst-case communication cost per client must be at least Ω(dlog(n²ε²/d)) to achieve O(d/(n²ε²)) centralized error, which matches the error under central DP. Despite this bound, we leverage the near-sparse structure of model updates, evidenced through recent empirical studies, to obtain improved tradeoffs for distributed \DP. In particular, we leverage linear compression methods, namely sketching, to attain compression rates of up to 50× with no significant decrease in model test accuracy achieving a noise multiplier 0.5. Our work demonstrates that fundamental tradeoffs in differentially private federated learning can be drastically improved in practice.

Dmitrii Usynin, Alexander Ziller, Moritz Knolle, Andrew Trask, Kritika Prakash, Daniel Rueckert, Georgios Kaissis
An automatic differentiation system for the age of differential privacy [arxiv] [Visit Poster at Spot A0]

We introduce Tritium, an automatic differentiation-based sensitivity analysis framework for differentially private (DP) machine learning (ML). Optimal noise calibration in this setting requires efficient Jacobian matrix computations and tight bounds on the L2-sensitivity. Our framework achieves these objectives by relying on a functional analysis-based method for sensitivity tracking, which we briefly outline. This approach interoperates naturally and seamlessly with static graph-based automatic differentiation, which enables order-of-magnitude improvements in compilation times compared to previous work. Moreover, we demonstrate that optimising the sensitivity of the entire computational graph at once yields substantially tighter estimates of the true sensitivity compared to interval bound propagation techniques. Our work naturally befits recent developments in DP such as individual privacy accounting, aiming to offer improved privacy-utility trade-offs, and represents a step towards the integration of accessible machine learning tooling with advanced privacy accounting systems.

Ashkan Yousefpour, Igor Shilov, Alexandre Sablayrolles, Davide Testuggine, Karthik Prasad, Mani Malek, John Nguyen, Sayan Ghosh, Akash Bharadwaj, Jessica Zhao, Graham Cormode, Ilya Mironov
Opacus: User-Friendly Differential Privacy Library in PyTorch [openreview] [Visit Poster at Spot D3]

We introduce Opacus, a free, open-source PyTorch library for training deep learning models with differential privacy (hosted at https://opacus.ai). Opacus is designed for simplicity, flexibility, and speed. It provides a simple and user-friendly API, and enables machine learning practitioners to make a training pipeline private by adding as little as two lines to their code. It supports a wide variety of layers, including multi-head attention, convolution, LSTM, and embedding, right out of the box, and it also provides the means for supporting other user-defined layers. Opacus computes batched per-sample gradients, providing better efficiency compared to the traditional “micro batch” approach. In this paper we present Opacus, detail the principles that drove its implementation and unique features, and compare its performance against other frameworks for differential privacy in ML.

Lennart Braun, Rosario Cammarota, Thomas Schneider
A Generic Hybrid 2PC Framework with Application to Private Inference of Unmodified Neural Networks (Extended Abstract) [openreview] [Visit Poster at Spot G3]

We present a new framework for generic mixed-protocol secure two-party computation (2PC) and private evaluation of neural networks based on the recent MOTION framework (Braun et al., ePrint '20). We implement five different 2PC protocols in the semi-honest setting -- Yao's garbled circuits, arithmetic and Boolean variants of Goldreich-Micali-Wigderson (GMW), and two secret-sharing-based protocols from ABY2.0 (Patra et al., USENIX Security '21) -- together with 20 conversions among each other and new optimizations. We explore the feasibility of evaluating neural networks with 2PC without making modifications to their structure, and provide secure tensor data types and specialized building blocks for common tensor operations. By supporting the Open Neural Network Exchange (ONNX) file format, this yields an easy-to-use solution for privately evaluating neural networks, and is interoperable with industry-standard deep learning frameworks such as TensorFlow and PyTorch. By exploiting the networks' high-level structure and using common 2PC techniques, we obtain a performance that is comparable to that of recent, highly optimized works and significantly better than when using generic 2PC for low-level hybrid circuits.

Jordan Awan, Vinayak Rao
Privacy-Aware Rejection Sampling (oral) [openreview] [Visit Poster at Spot B2]

Differential privacy (DP) offers strong protection against adversaries with arbitrary side-information and computational power. However, many implementations of DP mechanisms leave themselves vulnerable to side channel attacks, such as timing attacks. As many privacy mechanisms, such as the exponential mechanism, do not lend themselves to easy implementations, when sampling methods such as MCMC or rejection sampling are used, the runtime can leak privacy. In this work, we quantify the privacy cost due to the runtime of a rejection sampler in terms of (ε, δ)-DP. We also propose three modifications to the rejection sampling algorithm, to protect against timing attacks by making the runtime independent of the data. We also use our techniques to develop an adaptive rejection sampler for log-Holder densities, which also has data-independent runtime.

Jordan Awan, Salil Vadhan
Canonical Noise Distributions and Private Hypothesis Tests (oral) [openreview] [Visit Poster at Spot B3]

In the setting of f-DP, we propose the concept \emph{canonical noise distribution} (CND) which captures whether an additive privacy mechanism is tailored for a given f, and give a construction of a CND for an arbitrary tradeoff function . We show that private hypothesis tests are intimately related to CNDs, allowing for the release of private p-values at no additional privacy cost as well as the construction of uniformly most powerful (UMP) tests for binary data. We apply our techniques to difference of proportions testing.

Amir Mohammad Abouei, Clement Louis Canonne
Differential Privacy via Group Shuffling (oral) [openreview] [Visit Poster at Spot D2]

Xuechen Li, Florian Tramer, Percy Liang, Tatsunori Hashimoto
Simple Baselines Are Strong Performers for Differentially Private Natural Language Processing (oral) [openreview] [Visit Poster at Spot E1]

Differentially private learning has seen limited success for deep learning models of text, resulting in a perception that differential privacy may be incompatible with the language model fine-tuning paradigm. We demonstrate that this perception is inaccurate and that with the right setup, high performing private models can be learned on moderately-sized corpora by directly fine-tuning with differentially private optimization. Our work highlights the important role of hyperparameters, task formulations, and pretrained models. Our analyses also show that the low performance of naive differentially private baselines in prior work is attributable to suboptimal choices in these factors. Empirical results reveal that differentially private optimization does not suffer from dimension-dependent performance degradation with pretrained models and achieves performance on-par with state-of-the-art private training procedures and strong non-private baselines.

Robert Istvan Busa-Fekete, Andres Munoz medina, Umar Syed, Sergei Vassilvitskii
Population Level Privacy Leakage in Binary Classification wtih Label Noise (oral) [openreview] [Visit Poster at Spot F2]

We study the privacy limitations of label differential privacy. Label differential privacy has emerged as an intermediate trust model between local and central differential privacy, where only the label of each training example is protected (and the features are assumed to be public). We show that the guarantees provided by label DP are significantly weaker than they appear, as an adversary can "un-noise" the perturbed labels. Formally we show that the privacy loss has a close connection with Jeffreys' divergence of the conditional distribution between positive and negative labels, which allows explicit formulation of the trade-off between utility and privacy in this setting. Our results suggest how to select public features that optimize this trade-off. But we still show that there is no free lunch --- instances where label differential privacy guarantees are strong are exactly those where a good classifier does not exist. We complement the negative results with a non-parametric estimator for the true privacy loss, and apply our techniques on large-scale benchmark data to demonstrate how to achieve a desired privacy protection.

Aditya Hegde, Helen Möllering, Thomas Schneider, Hossein Yalame
SoK: Privacy-preserving Clustering (Extended Abstract) (oral) [openreview] [Visit Poster at Spot H1]

Block A Paris time
10:20-10:30	Welcome
10:30-11:00	Invited talk: Emiliano de Cristofaro (University College London)
Privacy in Machine Learning -- It's Complicated
11:00-11:15	Emiliano de Cristofaro Q & A:

11:15-11:30	Break
11:30-11:45	Differential Privacy via Group Shuffling (contributed talk) Amir Mohammad Abouei, Clement Louis Canonne
The past decade has seen data privacy emerge as a fundamental and pressing issue. Among the tools developed to tackle it, differential privacy has emerged as a central and principled framework, with specific variants capturing various threat models. In particular, the recently proposed shuffle model of differential privacy allows for promising tradeoffs between accuracy and privacy. However, the shuffle model may not be suitable in all situations, as it relies on a distributed setting where all users can coordinate and trust (or simulate) a joint shuffling algorithm. To address this, we introduce a new model, the group shuffle model, in which users are partitioned into several groups, each group having its own local shuffler. We investigate the privacy/accuracy tradeoffs in our model, by comparing it to both the shuffle and local models of privacy, which it some sense interpolates between. In addition to general relations between group shuffle, shuffle, and local privacy, we provide a detailed comparison of the cost and benefit of the group shuffle model, by providing both upper and lower bounds for the specific task of binary summation.
11:45-12:00	SoK: Privacy-preserving Clustering (Extended Abstract) (contributed talk) Aditya Hegde, Helen Möllering, Thomas Schneider, Hossein Yalame
Clustering is a popular unsupervised machine learning technique that groups similar input elements into clusters. In many applications, sensitive information is clustered that should not be leaked. Moreover, nowadays it is often required to combine data from multiple sources to increase the quality of the analysis as well as to outsource complex computation to powerful cloud servers. This calls for efficient privacy-preserving clustering. In this work, we systematically analyze the state-of-the-art in privacy-preserving clustering. We implement and benchmark today's four most efficient fully private clustering protocols by Cheon et al. (SAC'19), Meng et al. (ArXiv'19), Mohassel et al. (PETS'20), and Bozdemir et al. (ASIACCS'21) with respect to communication, computation, and clustering quality.
12:00-12：15	Contribute talk Q&A
12:15-12：30	Coffee Break
12:30-13:30	Poster
13:30:14:15	Panel
Block B LA time
8:20-8:30	Welcome
8:30-9:00	Invited talk: Helen Nissenbaum (Cornell Tech)
Practical Privacy, Fairness, Ethics, Policy
9:00-9:30	Invited talk: Aaron Roth (UPenn / Amazon)
Machine Unlearning.
9:30-10:00	Q & A for Helen and Aaron:
10:00-10:15	Break
10:15-11:15	Poster Session
11:15-11:30	Break
11:30-12:00	Invited talk: Kristin Lauter (Facebook AI Research):
ML on Encrypted Data.
12:00-12:15	Kristin Lauter Q & A :
12:15-12:30	Privacy-Aware Rejection Sampling (contributed talk) Jordan Awan, Vinayak Rao
Differential privacy (DP) offers strong protection against adversaries with arbitrary side-information and computational power. However, many implementations of DP mechanisms leave themselves vulnerable to side channel attacks, such as timing attacks. As many privacy mechanisms, such as the exponential mechanism, do not lend themselves to easy implementations, when sampling methods such as MCMC or rejection sampling are used, the runtime can leak privacy. In this work, we quantify the privacy cost due to the runtime of a rejection sampler in terms of -DP. We also propose three modifications to the rejection sampling algorithm, to protect against timing attacks by making the runtime independent of the data. We also use our techniques to develop an adaptive rejection sampler for log-Holder densities, which also has data-independent runtime.
12:30-12:45	Population Level Privacy Leakage in Binary Classification wtih Label Noise (contributed talk) Robert Istvan Busa-Fekete, Andres Munoz medina, Umar Syed, Sergei Vassilvitskii
We study the privacy limitations of label differential privacy. Label differential privacy has emerged as an intermediate trust model between local and central differential privacy, where only the label of each training example is protected (and the features are assumed to be public). We show that the guarantees provided by label DP are significantly weaker than they appear, as an adversary can "un-noise" the perturbed labels. Formally we show that the privacy loss has a close connection with Jeffreys' divergence of the conditional distribution between positive and negative labels, which allows explicit formulation of the trade-off between utility and privacy in this setting. Our results suggest how to select public features that optimize this trade-off. But we still show that there is no free lunch --- instances where label differential privacy guarantees are strong are exactly those where a good classifier does not exist. We complement the negative results with a non-parametric estimator for the true privacy loss, and apply our techniques on large-scale benchmark data to demonstrate how to achieve a desired privacy protection.
12:45-13:00	Simple Baselines Are Strong Performers for Differentially Private Natural Language Processing (contributed talk) Xuechen Li, Florian Tramer, Percy Liang, Tatsunori Hashimoto
Differentially private learning has seen limited success for deep learning models of text, resulting in a perception that differential privacy may be incompatible with the language model fine-tuning paradigm. We demonstrate that this perception is inaccurate and that with the right setup, high performing private models can be learned on moderately-sized corpora by directly fine-tuning with differentially private optimization. Our work highlights the important role of hyperparameters, task formulations, and pretrained models. Our analyses also show that the low performance of naive differentially private baselines in prior work is attributable to suboptimal choices in these factors. Empirical results reveal that differentially private optimization does not suffer from dimension-dependent performance degradation with pretrained models and achieves performance on-par with state-of-the-art private training procedures and strong non-private baselines.
13:00-13:15	Canonical Noise Distributions and Private Hypothesis Tests (contributed talk) Jordan Awan, Salil Vadhan
In the setting of -DP, we propose the concept \emph{canonical noise distribution} (CND) which captures whether an additive privacy mechanism is tailored for a given , and give a construction of a CND for an arbitrary tradeoff function . We show that private hypothesis tests are intimately related to CNDs, allowing for the release of private -values at no additional privacy cost as well as the construction of uniformly most powerful (UMP) tests for binary data. We apply our techniques to difference of proportions testing.
13:15-13：45	Q&A for four contributed talks
13:45-14:30	Panel
14:30-14：40	Closing

Privacy in Machine Learning

Scope

Call For Papers & Important Dates

Submission Instructions

Invited Speakers

Schedule

Accepted Papers

Organization

Workshop organizers

Program Committee