AI Hardware Security

Home Research Publications Projects Service Updates Blog CV

This direction focuses on the security and reliability of AI systems at the hardware and deployment boundary. I study bit-flip attacks on LLMs, accelerator-level fault assessment, approximate DNN fault behavior, and runtime methods for keeping edge inference safe under hardware-induced failures.

The goal is to evaluate and defend AI models as deployed systems, not just as abstract software models: faults, accelerators, memory behavior, approximate computing, thermal limits, and edge constraints all become part of the security problem.

Relevant Papers and Projects

FlipLLM: Efficient Bit-Flip Attacks on Multimodal LLMs using Reinforcement Learning
K. Khalil, K. A. Hoque
IEEE HOST 2026. Accepted.

RIFT: A Scalable Methodology for LLM Accelerator Fault Assessment using Reinforcement Learning
K. Khalil, K. A. Hoque
DATE 2026. Accepted. Best paper nomination.

EPSILON: Adaptive Fault Mitigation in Approximate Deep Neural Network using Statistical Signatures
K. Khalil, K. A. Hoque
IJCNN 2025.

VERMITHOR: Formally Verified Runtime Orchestration for Thermally-Safe Edge CPS Inference
K. Khalil, K. A. Hoque
Under review.

flipRL repository
Reinforcement-learning-driven bit-flip attack research.

Approximate Computing repository
Approximate DNN and hardware fault studies.