|
I am a PhD student in computer science (PhD CSE) at the Pennsylvania State University. I am advised by Prof. Gang Tan and work on developing secure and trustworthy software and machine learning systems. Currently, I am working on formal methods for trustworthy code translation and the security of LLM agents. In general, I am interested in topics in the intersection security, privacy, and machine learning. CV / Google Scholar / LinkedIn / Twitter / GitHub My Erdős number is 4: Paul Erdős → Israel Koren → Francesco Regazzoni → Takanori Isobe → Vishnu Asutosh Dasu I would love to collaborate with like-minded researchers and am open to opportunities. Please reach out to me if you're interested! |
|
|
|
|
|
|
|
|
Without special annotation (* represents equal contribution), authorship is in order of contribution as per the conventions of security and machine learning research. |
|
|
Md Rafi ur Rashid, Md Sadik Hossain Shanto, Vishnu Asutosh Dasu, Shagufta Mehnaz ACM Conference on Computer and Communications Security (CCS), 2026 Modern Vision-Language Models (VLMs) are robust to visual jailbreaks that use single, holistic images, but their safety alignment does not account for harmful semantics that are distributed across multiple image fragments. We introduce split-image visual jailbreak attacks (SIVA) that exploit this misalignment, evolving from naive splitting to an adaptive white-box attack and culminating in a black-box transfer attack powered by a novel adversarial knowledge distillation (Adv-KD) algorithm. Across four state-of-the-art VLMs and three jailbreak datasets, our strongest attack achieves up to 44% higher transfer success than existing baselines, and we propose efficient ways to address this vulnerability in VLM safety alignment. |
|
Runkai Zheng, Vishnu Asutosh Dasu, Yinong Oliver Wang, Haohan Wang, Fernando De la Torre International Conference on Computer Vision (ICCV), 2025 (Acceptance Rate: 24.0%) Dataset Distillation (DD) is an emerging technique that creates compact synthetic datasets to train models efficiently, but incorporating privacy through Differential Privacy (DP) often leads to degraded performance due to noisy optimization. This work introduces a novel framework that decouples sampling from optimization and improves signal quality by operating in an informative subspace, all without increasing the privacy cost. On CIFAR-10, it achieves a 10.0% accuracy gain with 50 images per class and an 8.3% improvement using just one-fifth the data compared to prior state-of-the-art private DD methods. |
|
Vishnu Asutosh Dasu, Md Rafi ur Rashid, Vipul Gupta, Saeid Tizpaz-Niari, Gang Tan International Conference on Software Engineering (ICSE) 2026 (Acceptance Rate: 22.4%) We explore pruning attention heads as a post-processing bias mitigation method for large language models (LLMs), addressing fairness concerns as LLMs expand into sensitive social contexts. Our Attention Pruning approach, using fairness-aware surrogate simulated annealing, achieves up to 40% reduction in gender bias and outperforms state-of-the-art bias mitigation strategies. |
|
Aydin Abadi*, Vishnu Asutosh Dasu*, Sumanta Sarkar* Network and Distributed System Security (NDSS) Symposium, 2025 (Acceptance Rate: 16.09%) We develop a novel variant of the Private Set Intersection (PSI) protocol that securely removes all pairwise duplicates among datasets held by two more users in federated learning. Our protocol preserves user privacy and does not reveal any information about the datasets to the users and server. The deduplication protocol results in significant improvement to LLM perplexity and reduces GPU training time. |
|
Vishnu Asutosh Dasu, Ashish Kumar, Saeid Tizpaz-Niari, Gang Tan The 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA), 2024 (Acceptance Rate: 20.6%) We develop randomized algorithms to mitigate unfairness in deep learning by dropping a subset of neurons from a trained neural network during at inference time. Our results show that a subset of neurons disparately contributes to unfairness and dropping them out during inference can improve fairness by up to 69%. |
|
Md Rafi ur Rashid, Vishnu Asutosh Dasu, Ye Wang, Gang Tan, Shagufta Mehnaz 40th Annual AAAI Conference on Artificial Intelligence (AAAI 2026) (Acceptance Rate: 22.65%, Selected for Oral Presentation) Large Language Models (LLMs) face a range of safety risks, but existing defenses tend to be narrow or overly rigid. We introduce Adversarial Scenario Extrapolation (ASE), an inference-time method that uses Chain-of-Thought reasoning to help LLMs anticipate adversarial scenarios and generate safer, more seamless responses. ASE outperforms six state-of-the-art defenses across multiple benchmarks, achieving near-zero jailbreak success rates, reduced bias, and fewer rejections. |
|
Md Rafi ur Rashid, Vishnu Asutosh Dasu, Kang Gu, Najrin Sultana, Shagufta Mehnaz 26th Privacy Enhancing Technologies Symposium (PETS), 2026 (Acceptance Rate: 20%) We introduce two novel privacy leakage attacks against federated language models. First, we show that intermediate model snapshots can leak more sensitive data than the final trained model. Second, we show that tampering with a model's selective weights responsible for memorizing sensitive data can aggravate privacy leakage. Our best-performing method outperforms existing attacks with stronger adversary assumptions. |
|
Alireza Aghabagherloo, Aydin Abadi, Sumanta Sarkar, Vishnu Asutosh Dasu, Bart Preneel IEEE Deep Learning Security and Privacy Workshop (DLSP), IEEE S&P 2025 (Acceptance Rate: 61.5%) We study the impact of duplicated images in training sets on image classifier performance, showing that duplication negatively affects training efficiency and accuracy, especially in adversarially trained models. The detrimental effect is more pronounced when duplication is uneven across classes, and increasing uniform duplication offers no significant accuracy improvement. |
|
Team EvoquerBOT, Penn State University Alexa Prize TaskBot Challenge 2 Proceedings, 2023 EvoquerBot is a multimedia chatbot developed for the TaskBot challenge, aimed at assisting users with cooking and DIY tasks in a single session. The bot addresses challenges like short development time, data quality, multimedia responses, and tailored conversation flow using agile classifier development, data augmentation, multimedia response design, and domain-specific dialogue state machines, ultimately improving user experience through superior task recommendations. |
|
Anubhab Baksi, Jakub Breier, Vishnu Asutosh Dasu, Xiaolu Hou, Hyunji Kim, Hwajeong Seo IEEE Access, 2023 (Acceptance Rate: 27.0%) ePrint We show new machine learning differential distinguishers for unkeyed and round-reduced versions of SPECK-32, SPECK-128, ASCON, SIMECK-32, SIMECK-64, and SKINNY-128. Our comprehensive experiments utilize neural networks and support vector machines in various settings and numerous input difference tuples. |
|
Vishnu Asutosh Dasu, Sumanta Sarkar, Kalikinkar Mandal ACM Workshop on Artificial Intelligence and Security (AISec), ACM CCS 2022 (Acceptance Rate: 35.0%) We propose PROV-FL, a secure and private federated learning protocol. PROV-FL utilizes homomorphic encryption and differential privacy to provide strong privacy guarantees. It is resilient to user dropouts/joins, supports verifiable aggregation, and requires only a single round of communication without a full-trusted third party. |
|
Satyam Kumar, Vishnu Asutosh Dasu, Anubhab Baksi, Santanu Sarkar, Dirmanto Jap, Jakub Breier, Shivam Bhasin IACR Conference on Cryptographic Hardware and Embedded Systems (CHES), 2022 (Acceptance Rate: 18.6%) code (Artifact Evaluated) We propose an end-to-end solution to perform SCA on stream ciphers by combining automated tools such as ML, MILP, and SMT. We demonstrate its efficacy by taking electromagnetic traces from a 32-bit software platform and performing SCA on the TRIVIUM stream cipher. |
|
Anubhab Baksi, Vishnu Asutosh Dasu, Banashri Karmakar, Anupam Chattopadhyay, Takanori Isobe INDOCRYPT, 2021 (Acceptance Rate: 41.5%) talk / ePrint / code We develop a method to extend the Boyar-Peralta's algorithm to use XOR3 gates, add XOR3 gates to existing XOR2 implementations, and show several SOTA results on the linear layers of block ciphers using different logic libraries. |
|
Anubhab Baksi, Banashri Karmakar, Vishnu Asutosh Dasu, Dhiman Saha, Anupam Chattopadhyay Security and Implementation of Lightweight Cryptography Workshop (SILC), EUROCRYPT 2021 We provide new insights for different notions of XOR count, develop methods to find sequential XOR count implementations using SMT and MILP, and present new results using XOR3 gates on the AES MixColumn matrix. |
|
Vishnu Asutosh Dasu, Anubhab Baksi, Sumanta Sarkar, Anupam Chattopadhyay IEEE International System-on-Chip Conference (SOCC), 2019 code We develop a framework that extends LIGHTER to add support for generating optimized implementations of 4x4 SBoxes using reversible logic libraries. |
|
Vishnu Asutosh Dasu, Midhush Manohar T.K. ReScience C, Volume 8, Issue 2, 2022 project page / code (Artifact Evaluated) / openreview / colab / blog We reproduce the results and validate the claims presented in GANSpace: Discovering Interpretable GAN Controls. |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Ritwik Sarkar* and Vishnu Asutosh Dasu* Presented at the Fifth Summer School on Computer Vision, Graphics, and Image Processing, Indian Statistical Institute (ISI), Kolkata, 2018 slides Recipient of the Best Project Award We develop a technique to determine the 3D coordinates of a human from a live video feed using a camera with a single lens (monocamera setup). |
|
|
|
|
I enjoy powerlifitng, playing the guitar, and reading about history, theology, and philosophy. |