arXiv subjects

Ashish Khisti

Publications and source records attributed to Ashish Khisti.

At least 37 records · Page 2Linked to original sources

arXiv2024

Secure Inference for Vertically Partitioned Data Using Multiparty Homomorphic Encryption

We propose a secure inference protocol for a distributed setting involving a single server node and multiple client nodes. We assume that the observed data vector is partitioned across multiple client nodes while the deep learning model is located at the server node. Each client node is required to encrypt its portion of the data vector and transmit the resulting ciphertext to the server node. The server node is required to collect the ciphertexts and perform inference in the encrypted domain. We demonstrate an application of multi-party homomorphic encryption (MPHE) to satisfy these requirements. We propose a packing scheme, that enables the server to form the ciphertext of the complete data by aggregating the ciphertext of data subsets encrypted using MPHE. While our proposed protocol builds upon prior horizontal federated training protocol~\cite{sav2020poseidon}, we focus on the inference for vertically partitioned data and avoid the transmission of (encrypted) model weights from the server node to the client nodes.

cs.CR↗

arXiv2024

Information Compression in the AI Era: Recent Advances and Future Challenges

This survey articles focuses on emerging connections between the fields of machine learning and data compression. While fundamental limits of classical (lossy) data compression are established using rate-distortion theory, the connections to machine learning have resulted in new theoretical analysis and application areas. We survey recent works on task-based and goal-oriented compression, the rate-distortion-perception theory and compression for estimation and inference. Deep learning based approaches also provide natural data-driven algorithmic approaches to compression. We survey recent works on applying deep learning techniques to task-based or goal-oriented compression, as well as image and video compression. We also discuss the potential use of large language models for text compression. We finally provide some directions for future research in this promising field.

cs.IT↗

arXiv2024

SECO: Secure Inference With Model Splitting Across Multi-Server Hierarchy

In the context of prediction-as-a-service, concerns about the privacy of the data and the model have been brought up and tackled via secure inference protocols. These protocols are built up by using single or multiple cryptographic tools designed under a variety of different security assumptions. In this paper, we introduce SECO, a secure inference protocol that enables a user holding an input data vector and multiple server nodes deployed with a split neural network model to collaboratively compute the prediction, without compromising either party's data privacy. We extend prior work on secure inference that requires the entire neural network model to be located on a single server node, to a multi-server hierarchy, where the user communicates to a gateway server node, which in turn communicates to remote server nodes. The inference task is split across the server nodes and must be performed over an encrypted copy of the data vector. We adopt multiparty homomorphic encryption and multiparty garbled circuit schemes, making the system secure against dishonest majority of semi-honest servers as well as protecting the partial model structure from the user. We evaluate SECO on multiple models, achieving the reduction of computation and communication cost for the user, making the protocol applicable to user's devices with limited resources.

cs.CR↗

arXiv2024

Importance Matching Lemma for Lossy Compression with Side Information

We propose two extensions to existing importance sampling based methods for lossy compression. First, we introduce an importance sampling based compression scheme that is a variant of ordered random coding (Theis and Ahmed, 2022) and is amenable to direct evaluation of the achievable compression rate for a finite number of samples. Our second and major contribution is the importance matching lemma, which is a finite proposal counterpart of the recently introduced Poisson matching lemma (Li and Anantharam, 2021). By integrating with deep learning, we provide a new coding scheme for distributed lossy compression with side information at the decoder. We demonstrate the effectiveness of the proposed scheme through experiments involving synthetic Gaussian sources, distributed image compression with MNIST and vertical federated learning with CIFAR-10.

cs.IT↗

arXiv2024

Subset Adaptive Relaying for Streaming Erasure Codes

This paper investigates adaptive streaming codes over a three-node relayed network. In this setting, a source transmits a sequence of message packets through a relay under a delay constraint of $T$ time slots per packet. The source-to-relay and relay-to-destination links are unreliable and introduce a maximum of $N_1$ and $N_2$ packet erasures respectively. Recent work has proposed adaptive (time variant) and nonadaptive (time invariant) code constructions for this setting and has shown that adaptive codes can achieve higher rates. However, the adaptive construction deals with many possibilities, leading to an impractical code with very large block lengths. In this work, we propose a simplified adaptive code construction which greatly improves the practicality of the code, with only a small cost to the achievable rates. We analyze the construction in terms of the achievable rates and field size requirements, and perform numerical simulations over statistical channels to estimate packet loss probabilities.

cs.IT↗

arXiv2024

On the Generalization of Stochastic Gradient Descent with Momentum

While momentum-based accelerated variants of stochastic gradient descent (SGD) are widely used when training machine learning models, there is little theoretical understanding on the generalization error of such methods. In this work, we first show that there exists a convex loss function for which the stability gap for multiple epochs of SGD with standard heavy-ball momentum (SGDM) becomes unbounded. Then, for smooth Lipschitz loss functions, we analyze a modified momentum-based update rule, i.e., SGD with early momentum (SGDEM) under a broad range of step-sizes, and show that it can train machine learning models for multiple epochs with a guarantee for generalization. Finally, for the special case of strongly convex loss functions, we find a range of momentum such that multiple epochs of standard SGDM, as a special form of SGDEM, also generalizes. Extending our results on generalization, we also develop an upper bound on the expected true risk, in terms of the number of training steps, sample size, and momentum. Our experimental evaluations verify the consistency between the numerical results and our theoretical bounds. SGDEM improves the generalization error of SGDM when training ResNet-18 on ImageNet in practical distributed settings.

cs.LG↗

arXiv2023

On the Choice of Perception Loss Function for Learned Video Compression

We study causal, low-latency, sequential video compression when the output is subjected to both a mean squared-error (MSE) distortion loss as well as a perception loss to target realism. Motivated by prior approaches, we consider two different perception loss functions (PLFs). The first, PLF-JD, considers the joint distribution (JD) of all the video frames up to the current one, while the second metric, PLF-FMD, considers the framewise marginal distributions (FMD) between the source and reconstruction. Using information theoretic analysis and deep-learning based experiments, we demonstrate that the choice of PLF can have a significant effect on the reconstruction, especially at low-bit rates. In particular, while the reconstruction based on PLF-JD can better preserve the temporal correlation across frames, it also imposes a significant penalty in distortion compared to PLF-FMD and further makes it more difficult to recover from errors made in the earlier output frames. Although the choice of PLF decisively affects reconstruction quality, we also demonstrate that it may not be essential to commit to a particular PLF during encoding and the choice of PLF can be delegated to the decoder. In particular, encoded representations generated by training a system to minimize the MSE (without requiring either PLF) can be {\em near universal} and can generate close to optimal reconstructions for either choice of PLF at the decoder. We validate our results using (one-shot) information-theoretic analysis, detailed study of the rate-distortion-perception tradeoff of the Gauss-Markov source model as well as deep-learning based experiments on moving MNIST and KTH datasets.

eess.IV↗

arXiv2023

Sequential Gradient Coding For Straggler Mitigation

In distributed computing, slower nodes (stragglers) usually become a bottleneck. Gradient Coding (GC), introduced by Tandon et al., is an efficient technique that uses principles of error-correcting codes to distribute gradient computation in the presence of stragglers. In this paper, we consider the distributed computation of a sequence of gradients $\{g(1),g(2),\ldots,g(J)\}$, where processing of each gradient $g(t)$ starts in round-$t$ and finishes by round-$(t+T)$. Here $T\geq 0$ denotes a delay parameter. For the GC scheme, coding is only across computing nodes and this results in a solution where $T=0$. On the other hand, having $T>0$ allows for designing schemes which exploit the temporal dimension as well. In this work, we propose two schemes that demonstrate improved performance compared to GC. Our first scheme combines GC with selective repetition of previously unfinished tasks and achieves improved straggler mitigation. In our second scheme, which constitutes our main contribution, we apply GC to a subset of the tasks and repetition for the remainder of the tasks. We then multiplex these two classes of tasks across workers and rounds in an adaptive manner, based on past straggler patterns. Using theoretical analysis, we demonstrate that our second scheme achieves significant reduction in the computational load. In our experiments, we study a practical setting of concurrently training multiple neural networks over an AWS Lambda cluster involving 256 worker nodes, where our framework naturally applies. We demonstrate that the latter scheme can yield a 16\% improvement in runtime over the baseline GC scheme, in the presence of naturally occurring, non-simulated stragglers.

cs.LG↗

arXiv2023

Quadratic Functional Encryption for Secure Training in Vertical Federated Learning

Vertical federated learning (VFL) enables the collaborative training of machine learning (ML) models in settings where the data is distributed amongst multiple parties who wish to protect the privacy of their individual data. Notably, in VFL, the labels are available to a single party and the complete feature set is formed only when data from all parties is combined. Recently, Xu et al. proposed a new framework called FedV for secure gradient computation for VFL using multi-input functional encryption. In this work, we explain how some of the information leakage in Xu et al. can be avoided by using Quadratic functional encryption when training generalized linear models for vertical federated learning.

cs.CR↗

arXiv2023

Random Edge Coding: One-Shot Bits-Back Coding of Large Labeled Graphs

We present a one-shot method for compressing large labeled graphs called Random Edge Coding. When paired with a parameter-free model based on Pólya's Urn, the worst-case computational and memory complexities scale quasi-linearly and linearly with the number of observed edges, making it efficient on sparse graphs, and requires only integer arithmetic. Key to our method is bits-back coding, which is used to sample edges and vertices without replacement from the edge-list in a way that preserves the structure of the graph. Optimality is proven under a class of random graph models that are invariant to permutations of the edges and of vertices within an edge. Experiments indicate Random Edge Coding can achieve competitive compression performance on real-world network datasets and scales to graphs with millions of nodes and edges.

cs.LG↗

arXiv2023

Compressing Multisets with Large Alphabets using Bits-Back Coding

Current methods which compress multisets at an optimal rate have computational complexity that scales linearly with alphabet size, making them too slow to be practical in many real-world settings. We show how to convert a compression algorithm for sequences into one for multisets, in exchange for an additional complexity term that is quasi-linear in sequence length. This allows us to compress multisets of exchangeable symbols at an optimal rate, with computational complexity decoupled from the alphabet size. The key insight is to avoid encoding the multiset directly, and instead compress a proxy sequence, using a technique called `bits-back coding'. We demonstrate the method experimentally on tasks which are intractable with previous optimal-rate methods: compression of multisets of images and JavaScript Object Notation (JSON) files. Code for our experiments is available at https://github.com/facebookresearch/multiset-compression.

cs.IT↗

arXiv2022

Optimal Streaming Erasure Codes over the Three-Node Relay Network

This paper investigates low-latency streaming codes for a three-node relay network. The source transmits a sequence of messages (streaming messages) to the destination through the relay between them, where the first-hop channel from the source to the relay and the second-hop channel from the relay to the destination are subject to packet erasures. Every source message must be recovered perfectly at the destination subject to a fixed decoding delay of $T$ time slots. In any sliding window of $T+1$ time slots, we assume no more than $N_1$ and $N_2$ erasures are introduced by the first-hop channel and second-hop channel respectively. Under this channel loss assumption, we fully characterize the maximum achievable rate in terms of $T$, $N_1$ and $N_2$. The achievability is proved by using a symbol-wise decode-forward strategy where the source symbols within the same message are decoded by the relay with different delays. The converse is proved by analyzing the maximum achievable rate for each channel when the erasures in the other channel are consecutive (bursty). In addition, we show that traditional message-wise decode-forward strategies, which require the source symbols within the same message to be decoded by the relay with the same delay, are sub-optimal in general.

cs.IT↗

arXiv2022

Adaptive relaying for streaming erasure codes in a three node relay network

This paper investigates adaptive streaming codes over a three-node relayed network. In this setting, a source node transmits a sequence of message packets to a destination through a relay. The source-to-relay and relay-to-destination links are unreliable and introduce at most $N_1$ and $N_2$ packet erasures, respectively. The destination node must recover each message packet within a strict delay constraint $T$. The paper presents achievable streaming codes for all feasible parameters $\{N_1, N_2, T\}$ that exploit the fact that the relay naturally observes the erasure pattern occurring in the link from source to relay, thus it can adapt its relaying strategy based on these observations. In a recent work, Fong et al. provide streaming codes featuring channel-state-independent relaying strategies. The codes proposed in this paper achieve rates higher than the ones proposed by Fong et al. whenever $N_2 > N_1$, and achieve the same rate when $N_2 = N_1$. The paper also presents an upper bound on the achievable rate that takes into account erasures in both links in order to bound the rate in the second link. The upper bound is shown to be tighter than a trivial bound that considers only the erasures in the second link.

cs.IT↗

arXiv2022

Variational Model Inversion Attacks

Given the ubiquity of deep neural networks, it is important that these models do not reveal information about sensitive data that they have been trained on. In model inversion attacks, a malicious user attempts to recover the private dataset used to train a supervised neural network. A successful model inversion attack should generate realistic and diverse samples that accurately describe each of the classes in the private dataset. In this work, we provide a probabilistic interpretation of model inversion attacks, and formulate a variational objective that accounts for both diversity and accuracy. In order to optimize this variational objective, we choose a variational family defined in the code space of a deep generative model, trained on a public auxiliary dataset that shares some structural similarity with the target dataset. Empirically, our method substantially improves performance in terms of target attack accuracy, sample realism, and diversity on datasets of faces and chest X-ray images.

cs.LG↗

arXiv2022

Error-correcting codes for low latency streaming over multiple link relay networks

This paper investigates the performance of streaming codes in low-latency applications over a multi-link three-node relayed network. The source wishes to transmit a sequence of messages to the destination through a relay. Each message must be reconstructed after a fixed decoding delay. The special case with one link connecting each node has been studied by Fong et. al [1], and a multi-hop multi-link setting has been studied by Domanovitz et. al [2]. The topology with three nodes and multiple links is studied in this paper. Each link is subject to a different number of erasures due to different channel conditions. An information-theoretic upper bound is derived, and an achievable scheme is presented. The proposed scheme judiciously allocates rates for each link based on the concept of delay spectrum. The achievable scheme is compared to two baseline schemes and the scheme proposed in [2]. Experimental results show that this scheme achieves higher rates than the other schemes, and can achieve the upper bound even in non-trivial scenarios. The scheme is further extended to handle different propagation delays in each link, something not previously considered in the literature. Simulations over statistical channels show that the proposed scheme can outperform the simpler baseline under practical models.

cs.IT↗

arXiv2021

Universal Rate-Distortion-Perception Representations for Lossy Compression

In the context of lossy compression, Blau & Michaeli (2019) adopt a mathematical notion of perceptual quality and define the information rate-distortion-perception function, generalizing the classical rate-distortion tradeoff. We consider the notion of universal representations in which one may fix an encoder and vary the decoder to achieve any point within a collection of distortion and perception constraints. We prove that the corresponding information-theoretic universal rate-distortion-perception function is operationally achievable in an approximate sense. Under MSE distortion, we show that the entire distortion-perception tradeoff of a Gaussian source can be achieved by a single encoder of the same rate asymptotically. We then characterize the achievable distortion-perception region for a fixed representation in the case of arbitrary distributions, identify conditions under which the aforementioned results continue to hold approximately, and study the case when the rate is not fixed in advance. This motivates the study of practical constructions that are approximately universal across the RDP tradeoff, thereby alleviating the need to design a new encoder for each objective. We provide experimental results on MNIST and SVHN suggesting that on image compression tasks, the operational tradeoffs achieved by machine learning models with a fixed encoder suffer only a small penalty when compared to their variable encoder counterparts.

cs.IT↗

arXiv2021

Streaming Erasure Codes over Multi-Access Relayed Networks

Many emerging multimedia streaming applications involve multiple users communicating under strict latency constraints. In this paper we study streaming codes for a network involving two source nodes, one relay node and a destination node. In our setting, each source node transmits a stream of messages, through the relay, to a destination, who is required to decode the messages under a strict delay constraint. For the case of a single source node, a class of streaming codes has been proposed by Fong et al. [2], using the concept of {\em delay-spectrum}. In the present work we present an in-depth analysis of the properties of delay-spectrum and apply them to develop streaming codes for our proposed setting through a novel framework. Our first scheme involves greedily selecting the rate on the link from relay to destination and using properties of the delay-spectrum to find feasible streaming codes that satisfy the required delay constraints. We provide a closed form expression for the achievable rate region and identify conditions when the proposed scheme is optimal by establishing a natural outer bound. Our second scheme builds upon this approach, but uses a numerical optimization-based approach to improve the achievable rate region over the first scheme. We demonstrate that our proposed schemes achieve significant improvements over baseline schemes based on single-user codes.

cs.IT↗

arXiv2021

On the Generalization of Stochastic Gradient Descent with Momentum

While momentum-based methods, in conjunction with stochastic gradient descent (SGD), are widely used when training machine learning models, there is little theoretical understanding on the generalization error of such methods. In this work, we first show that there exists a convex loss function for which algorithmic stability fails to establish generalization guarantees when SGD with standard heavy-ball momentum (SGDM) is run for multiple epochs. Then, for smooth Lipschitz loss functions, we analyze a modified momentum-based update rule, i.e., SGD with early momentum (SGDEM), and show that it admits an upper-bound on the generalization error. Thus, our results show that machine learning models can be trained for multiple epochs of SGDEM with a guarantee for generalization. Finally, for the special case of strongly convex loss functions, we find a range of momentum such that multiple epochs of standard SGDM, as a special form of SGDEM, also generalizes. Extending our results on generalization, we also develop an upper-bound on the expected true risk, in terms of the number of training steps, the size of the training set, and the momentum parameter. Experimental evaluations verify the consistency between the numerical results and our theoretical bounds and the effectiveness of SGDEM for smooth Lipschitz loss functions.

cs.LG↗