Dipkamal Bhusal, Michael Clifford, Sara Rampazzi, Nidhi Rastogi

40th Annual Conference on Neural Information Processing Systems (NeurIPS) 2025

In this work, we propose FACE (Faithful Automatic Concept Extraction), a novel framework that augments Non-negative Matrix Factorization (NMF) with a Kullback-Leibler (KL) divergence regularization term to ensure alignment between the model’s original and concept-based predictions. Unlike prior methods that operate solely on encoder activations, FACE incorporates classifier supervision during concept learning, enforcing predictive consistency and enabling faithful explanations.

Ryan L. Yang, Dipkamal Bhusal (Mentorship), Nidhi Rastogi

First Workshop on CogInterp: Interpreting Cognition in Deep Learning Models (NeurIPS 2025)

In this work, I mentored Ryan Yang, an undergraduate from Brown University who joined RIT for summer research. This project tackles a core challenge in deep learning interpretability, ensuring models are "right for the right reasons", by introducing a scalable framework that uses a vision-language model to extract semantic attention, training the model to ground its predictions in the correct evidence.

Ayushi Mehrotra, Derek Peng, Dipkamal Bhusal (Mentorship), Nidhi Rastogi

Workshop: Reliable ML from Unreliable Data. (NeurIPS 2025)

In this work, I mentored Ayushi and Derek, to design a patch agnostic defense against adversarial patch attack that leverages concept-based explanations to identify and suppress the most influential concept activation vectors, thereby neutralizing patch effects without explicit detection. Our method achieves higher robust and clean accuracy than the state-of-the-art PatchCleanser.

Sanish Suwal, Dipkamal Bhusal (Mentorship), Michael Clifford, Nidhi Rastogi

Workshop: First Workshop on CogInterp: Interpreting Cognition in Deep Learning Models. (NeurIPS 2025)

In this work, I mentored Sanish where we investigated how magnitude-based pruning followed by fine-tuning affects both low-level saliency maps and high-level concept representations.

Ayushi Mehrotra, Dipkamal Bhusal (Mentorship), Nidhi Rastogi

2nd Workshop on Attributing Model Behavior at Scale at NeurIPS 2024

This work introduces Hessian Sets, a technique that leverages the Hessian matrix to detect and attribute pairwise feature interactions in image classifiers. By integrating segmentation masks from the Segment Anything Model (SAM), we provide more interpretable and concise explanations. This was my first paper as a mentor. We are extending the work currently.

Dipkamal Bhusal, Md Tanvirul Alam, Le Nguyen, Ashim Mahara, Zachary Lightcap, Rodney, Romy, Grace, Benjamin, Nidhi Rastogi

40th Annual Computer Security Applications Conference (ACSAC) 2024

Large Language Models (LLMs) have demonstrated potential in cybersecurity applications but suffer from hallucinations and a lack of truthfulness. Existing benchmarks provide general evaluations but do not sufficiently address the practical and applied aspects of cybersecurity-specific tasks. To address this gap, we introduce the SECURE (Security Extraction, Understanding & Reasoning Evaluation), a benchmark designed to assess LLMs performance in realistic cybersecurity scenarios.

Md Tanvirul Alam*, Dipkamal Bhusal*, Le Nguyen, Nidhi Rastogi (* Equal Contribution)

39th Annual Conference on Neural Information Processing Systems (NeurIPS) 2024

We extend the knowledge intensive LLM evaluation framework proposed in SECURE and evaluate LLMs in CTI-specific tasks. Cyber threat intelligence (CTI) is crucial in today’s cybersecurity landscape, providing essential insights to understand and mitigate the ever-evolving cyber threats. CTIBench is a benchmark designed to assess LLMs’ performance in CTI applications.

Dipkamal Bhusal, Md Tanvirul Alam, Monish K. , Michael Clifford, Sara Rampazzi, Nidhi Rastogi

9th IEEE European Symposium on Security and Privacy (EuroS&P) 2024

We develop a practical method for utilizing sensitivity of model prediction and feature attribution to detect adversarial samples. Our method, PASA, requires the computation of two test statistics using model prediction and feature attribution and can reliably detect adversarial samples using thresholds learned from benign samples.

Dipkamal Bhusal, Rosalyn Shin, Ajay Ashok Shewale, Monish Kumar Manikya Veerabhadran, Michael Clifford, Sara Rampazzi, Nidhi Rastogi

18th International Conference on Availability, Reliability and Security(ARES) 2023

This paper provides a comprehensive analysis of explainable methods and demonstrates their efficacy in three distinct security applications: anomaly detection using system logs, malware prediction, and detection of adversarial images. Our quantitative and qualitative analysis reveals serious limitations and concerns in state-of-the-art explanation methods in all three applications.