arXiv subjects

Virat Shejwalkar

Publications and source records attributed to Virat Shejwalkar.

14 recordsLinked to original sources

arXiv2026

Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs

Fine-tuning on benign data is known to degrade safety alignment in text and vision LLMs, but whether distinct input properties drive this vulnerability differently remains unclear. Audio introduces a richer problem where benign samples can neighbor harmful content through what is said or how it sounds. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based framework that decomposes embedding-space distance into semantic, acoustic, and mixed axes. We find that the dominant vulnerability axis is architecture-conditioned, determined by how each model's encoder and projector transform audio into the backbone LLM's input space. Across three models, benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87%, with the most damaging axis shifting from semantic to acoustic proximity depending on encoder design. Mechanistically, fine-tuning selectively suppresses late-layer refusal circuits while frozen encoders preserve upstream representations: the model still detects harmful content but stops refusing, a recognition-refusal dissociation. Two practical defenses, filtering training data to maximize distance from harmful embeddings and a textual system prompt at inference, reduce JSR to near-zero without architectural modification. These findings show that safety evaluation should account for modality and architecture, while highlighting Audio LLMs as a useful testbed for understanding alignment fragility.

cs.CR↗

arXiv2025

Rebellion: Noise-Robust Reasoning Training for Audio Reasoning Models

Instilling reasoning capabilities in large models (LMs) using reasoning training (RT) significantly improves LMs' performances. Thus Audio Reasoning Models (ARMs), i.e., audio LMs that can reason, are becoming increasingly popular. However, no work has studied the safety of ARMs against jailbreak attacks that aim to elicit harmful responses from target models. To this end, first, we show that standard RT with appropriate safety reasoning data can protect ARMs from vanilla audio jailbreaks, but cannot protect them against our proposed simple yet effective jailbreaks. We show that this is because of the significant representation drift between vanilla and advanced jailbreaks which forces the target ARMs to emit harmful responses. Based on this observation, we propose Rebellion, a robust RT that trains ARMs to be robust to the worst-case representation drift. All our results are on Qwen2-Audio; they demonstrate that Rebellion: 1) can protect against advanced audio jailbreaks without compromising performance on benign tasks, and 2) significantly improves accuracy-safety trade-off over standard RT method.

cs.AI↗

arXiv2025

Multilingual and Multi-Accent Jailbreaking of Audio LLMs

Large Audio Language Models (LALMs) have significantly advanced audio understanding but introduce critical security risks, particularly through audio jailbreaks. While prior work has focused on English-centric attacks, we expose a far more severe vulnerability: adversarial multilingual and multi-accent audio jailbreaks, where linguistic and acoustic variations dramatically amplify attack success. In this paper, we introduce Multi-AudioJail, the first systematic framework to exploit these vulnerabilities through (1) a novel dataset of adversarially perturbed multilingual/multi-accent audio jailbreaking prompts, and (2) a hierarchical evaluation pipeline revealing that how acoustic perturbations (e.g., reverberation, echo, and whisper effects) interacts with cross-lingual phonetics to cause jailbreak success rates (JSRs) to surge by up to +57.25 percentage points (e.g., reverberated Kenyan-accented attack on MERaLiON). Crucially, our work further reveals that multimodal LLMs are inherently more vulnerable than unimodal systems: attackers need only exploit the weakest link (e.g., non-English audio inputs) to compromise the entire model, which we empirically show by multilingual audio-only attacks achieving 3.1x higher success rates than text-only attacks. We plan to release our dataset to spur research into cross-modal defenses, urging the community to address this expanding attack surface in multimodality as LALMs evolve.

cs.SD↗

arXiv2025

Decoding FL Defenses: Systemization, Pitfalls, and Remedies

While the community has designed various defenses to counter the threat of poisoning attacks in Federated Learning (FL), there are no guidelines for evaluating these defenses. These defenses are prone to subtle pitfalls in their experimental setups that lead to a false sense of security, rendering them unsuitable for practical deployment. In this paper, we systematically understand, identify, and provide a better approach to address these challenges. First, we design a comprehensive systemization of FL defenses along three dimensions: i) how client updates are processed, ii) what the server knows, and iii) at what stage the defense is applied. Next, we thoroughly survey 50 top-tier defense papers and identify the commonly used components in their evaluation setups. Based on this survey, we uncover six distinct pitfalls and study their prevalence. For example, we discover that around 30% of these works solely use the intrinsically robust MNIST dataset, and 40% employ simplistic attacks, which may inadvertently portray their defense as robust. Using three representative defenses as case studies, we perform a critical reevaluation to study the impact of the identified pitfalls and show how they lead to incorrect conclusions about robustness. We provide actionable recommendations to help researchers overcome each pitfall.

cs.CR↗

arXiv2024

Recycling Scraps: Improving Private Learning by Leveraging Intermediate Checkpoints

In this work, we focus on improving the accuracy-variance trade-off for state-of-the-art differentially private machine learning (DP ML) methods. First, we design a general framework that uses aggregates of intermediate checkpoints \emph{during training} to increase the accuracy of DP ML techniques. Specifically, we demonstrate that training over aggregates can provide significant gains in prediction accuracy over the existing state-of-the-art for StackOverflow, CIFAR10 and CIFAR100 datasets. For instance, we improve the state-of-the-art DP StackOverflow accuracies to 22.74\% (+2.06\% relative) for $ε=8.2$, and 23.90\% (+2.09\%) for $ε=18.9$. Furthermore, these gains magnify in settings with periodically varying training data distributions. We also demonstrate that our methods achieve relative improvements of 0.54\% and 62.6\% in terms of utility and variance, on a proprietary, production-grade pCVR task. Lastly, we initiate an exploration into estimating the uncertainty (variance) that DP noise adds in the predictions of DP ML models. We prove that, under standard assumptions on the loss function, the sample variance from last few checkpoints provides a good approximation of the variance of the final model of a DP run. Empirically, we show that the last few checkpoints can provide a reasonable lower bound for the variance of a converged DP model. Crucially, all the methods proposed in this paper operate on \emph{a single training run} of the DP ML technique, thus incurring no additional privacy cost.

cs.LG↗

arXiv2022

Security Analysis of SplitFed Learning

Split Learning (SL) and Federated Learning (FL) are two prominent distributed collaborative learning techniques that maintain data privacy by allowing clients to never share their private data with other clients and servers, and fined extensive IoT applications in smart healthcare, smart cities, and smart industry. Prior work has extensively explored the security vulnerabilities of FL in the form of poisoning attacks. To mitigate the effect of these attacks, several defenses have also been proposed. Recently, a hybrid of both learning techniques has emerged (commonly known as SplitFed) that capitalizes on their advantages (fast training) and eliminates their intrinsic disadvantages (centralized model updates). In this paper, we perform the first ever empirical analysis of SplitFed's robustness to strong model poisoning attacks. We observe that the model updates in SplitFed have significantly smaller dimensionality as compared to FL that is known to have the curse of dimensionality. We show that large models that have higher dimensionality are more susceptible to privacy and security attacks, whereas the clients in SplitFed do not have the complete model and have lower dimensionality, making them more robust to existing model poisoning attacks. Our results show that the accuracy reduction due to the model poisoning attack is 5x lower for SplitFed compared to FL.

cs.LG↗

arXiv2022

The Perils of Learning From Unlabeled Data: Backdoor Attacks on Semi-supervised Learning

Semi-supervised machine learning (SSL) is gaining popularity as it reduces the cost of training ML models. It does so by using very small amounts of (expensive, well-inspected) labeled data and large amounts of (cheap, non-inspected) unlabeled data. SSL has shown comparable or even superior performances compared to conventional fully-supervised ML techniques. In this paper, we show that the key feature of SSL that it can learn from (non-inspected) unlabeled data exposes SSL to strong poisoning attacks. In fact, we argue that, due to its reliance on non-inspected unlabeled data, poisoning is a much more severe problem in SSL than in conventional fully-supervised ML. Specifically, we design a backdoor poisoning attack on SSL that can be conducted by a weak adversary with no knowledge of target SSL pipeline. This is unlike prior poisoning attacks in fully-supervised settings that assume strong adversaries with practically-unrealistic capabilities. We show that by poisoning only 0.2% of the unlabeled training data, our attack can cause misclassification of more than 80% of test inputs (when they contain the adversary's backdoor trigger). Our attacks remain effective across twenty combinations of benchmark datasets and SSL algorithms, and even circumvent the state-of-the-art defenses against backdoor attacks. Our work raises significant concerns about the practical utility of existing SSL algorithms.

cs.CR↗

arXiv2022

FRL: Federated Rank Learning

Federated learning (FL) allows mutually untrusted clients to collaboratively train a common machine learning model without sharing their private/proprietary training data among each other. FL is unfortunately susceptible to poisoning by malicious clients who aim to hamper the accuracy of the commonly trained model through sending malicious model updates during FL's training process. We argue that the key factor to the success of poisoning attacks against existing FL systems is the large space of model updates available to the clients, allowing malicious clients to search for the most poisonous model updates, e.g., by solving an optimization problem. To address this, we propose Federated Rank Learning (FRL). FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values). To be able to train the global model using parameter ranks (instead of parameter weights), FRL leverage ideas from recent supermasks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch to generate the global ranking of the next training epoch. Intuitively, our voting-based aggregation mechanism prevents poisoning clients from making significant adversarial modifications to the global model, as each client will have a single vote! We demonstrate the robustness of FRL to poisoning through analytical proofs and experimentation. We also show FRL's high communication efficiency. Our experiments demonstrate the superiority of FRL in real-world FL settings.

cs.LG↗

arXiv2022

GECKO: Reconciling Privacy, Accuracy and Efficiency in Embedded Deep Learning

Embedded systems demand on-device processing of data using Neural Networks (NNs) while conforming to the memory, power and computation constraints, leading to an efficiency and accuracy tradeoff. To bring NNs to edge devices, several optimizations such as model compression through pruning, quantization, and off-the-shelf architectures with efficient design have been extensively adopted. These algorithms when deployed to real world sensitive applications, requires to resist inference attacks to protect privacy of users training data. However, resistance against inference attacks is not accounted for designing NN models for IoT. In this work, we analyse the three-dimensional privacy-accuracy-efficiency tradeoff in NNs for IoT devices and propose Gecko training methodology where we explicitly add resistance to private inferences as a design objective. We optimize the inference-time memory, computation, and power constraints of embedded devices as a criterion for designing NN architecture while also preserving privacy. We choose quantization as design choice for highly efficient and private models. This choice is driven by the observation that compressed models leak more information compared to baseline models while off-the-shelf efficient architectures indicate poor efficiency and privacy tradeoff. We show that models trained using Gecko methodology are comparable to prior defences against black-box membership attacks in terms of accuracy and privacy while providing efficiency.

cs.CR↗

arXiv2021

Back to the Drawing Board: A Critical Evaluation of Poisoning Attacks on Production Federated Learning

While recent works have indicated that federated learning (FL) may be vulnerable to poisoning attacks by compromised clients, their real impact on production FL systems is not fully understood. In this work, we aim to develop a comprehensive systemization for poisoning attacks on FL by enumerating all possible threat models, variations of poisoning, and adversary capabilities. We specifically put our focus on untargeted poisoning attacks, as we argue that they are significantly relevant to production FL deployments. We present a critical analysis of untargeted poisoning attacks under practical, production FL environments by carefully characterizing the set of realistic threat models and adversarial capabilities. Our findings are rather surprising: contrary to the established belief, we show that FL is highly robust in practice even when using simple, low-cost defenses. We go even further and propose novel, state-of-the-art data and model poisoning attacks, and show via an extensive set of experiments across three benchmark datasets how (in)effective poisoning attacks are in the presence of simple defense mechanisms. We aim to correct previous misconceptions and offer concrete guidelines to conduct more accurate (and more realistic) research on this topic.

cs.LG↗

arXiv2021

Mitigating Membership Inference Attacks by Self-Distillation Through a Novel Ensemble Architecture

Membership inference attacks are a key measure to evaluate privacy leakage in machine learning (ML) models. These attacks aim to distinguish training members from non-members by exploiting differential behavior of the models on member and non-member inputs. The goal of this work is to train ML models that have high membership privacy while largely preserving their utility; we therefore aim for an empirical membership privacy guarantee as opposed to the provable privacy guarantees provided by techniques like differential privacy, as such techniques are shown to deteriorate model utility. Specifically, we propose a new framework to train privacy-preserving models that induces similar behavior on member and non-member inputs to mitigate membership inference attacks. Our framework, called SELENA, has two major components. The first component and the core of our defense is a novel ensemble architecture for training. This architecture, which we call Split-AI, splits the training data into random subsets, and trains a model on each subset of the data. We use an adaptive inference strategy at test time: our ensemble architecture aggregates the outputs of only those models that did not contain the input sample in their training data. 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. The Self-Distillation component (self-)distills the training dataset through our Split-AI ensemble, without using any external public datasets. Through extensive experiments on major benchmark datasets we show that SELENA presents a superior trade-off between membership privacy and utility compared to the state of the art.

cs.CR↗

arXiv2021

Quantifying Privacy Leakage in Graph Embedding

Graph embeddings have been proposed to map graph data to low dimensional space for downstream processing (e.g., node classification or link prediction). With the increasing collection of personal data, graph embeddings can be trained on private and sensitive data. For the first time, we quantify the privacy leakage in graph embeddings through three inference attacks targeting Graph Neural Networks. We propose a membership inference attack to infer whether a graph node corresponding to individual user's data was member of the model's training or not. We consider a blackbox setting where the adversary exploits the output prediction scores, and a whitebox setting where the adversary has also access to the released node embeddings. This attack provides an accuracy up to 28% (blackbox) 36% (whitebox) beyond random guess by exploiting the distinguishable footprint between train and test data records left by the graph embedding. We propose a Graph Reconstruction attack where the adversary aims to reconstruct the target graph given the corresponding graph embeddings. Here, the adversary can reconstruct the graph with more than 80% of accuracy and link inference between two nodes around 30% more confidence than a random guess. We then propose an attribute inference attack where the adversary aims to infer a sensitive attribute. We show that graph embeddings are strongly correlated to node attributes letting the adversary inferring sensitive information (e.g., gender or location).

cs.CR↗

arXiv2020

Membership Privacy for Machine Learning Models Through Knowledge Transfer

Large capacity machine learning (ML) models are prone to membership inference attacks (MIAs), which aim to infer whether the target sample is a member of the target model's training dataset. The serious privacy concerns due to the membership inference have motivated multiple defenses against MIAs, e.g., differential privacy and adversarial regularization. Unfortunately, these defenses produce ML models with unacceptably low classification performances. Our work proposes a new defense, called distillation for membership privacy (DMP), against MIAs that preserves the utility of the resulting models significantly better than prior defenses. DMP leverages knowledge distillation to train ML models with membership privacy. We provide a novel criterion to tune the data used for knowledge transfer in order to amplify the membership privacy of DMP. Our extensive evaluation shows that DMP provides significantly better tradeoffs between membership privacy and classification accuracies compared to state-of-the-art MIA defenses. For instance, DMP achieves ~100% accuracy improvement over adversarial regularization for DenseNet trained on CIFAR100, for similar membership privacy (measured using MIA risk): when the MIA risk is 53.7%, adversarially regularized DenseNet is 33.6% accurate, while DMP-trained DenseNet is 65.3% accurate.

cs.LG↗

arXiv2019

Cronus: Robust and Heterogeneous Collaborative Learning with Black-Box Knowledge Transfer

Collaborative (federated) learning enables multiple parties to train a model without sharing their private data, but through repeated sharing of the parameters of their local models. Despite its advantages, this approach has many known privacy and security weaknesses and performance overhead, in addition to being limited only to models with homogeneous architectures. Shared parameters leak a significant amount of information about the local (and supposedly private) datasets. Besides, federated learning is severely vulnerable to poisoning attacks, where some participants can adversarially influence the aggregate parameters. Large models, with high dimensional parameter vectors, are in particular highly susceptible to privacy and security attacks: curse of dimensionality in federated learning. We argue that sharing parameters is the most naive way of information exchange in collaborative learning, as they open all the internal state of the model to inference attacks, and maximize the model's malleability by stealthy poisoning attacks. We propose Cronus, a robust collaborative machine learning framework. The simple yet effective idea behind designing Cronus is to control, unify, and significantly reduce the dimensions of the exchanged information between parties, through robust knowledge transfer between their black-box local models. We evaluate all existing federated learning algorithms against poisoning attacks, and we show that Cronus is the only secure method, due to its tight robustness guarantee. Treating local models as black-box, reduces the information leakage through models, and enables us using existing privacy-preserving algorithms that mitigate the risk of information leakage through the model's output (predictions). Cronus also has a significantly lower sample complexity, compared to federated learning, which does not bind its security to the number of participants.

stat.ML↗