Research Projects

Below are some of my key research projects in machine learning, AI security, and hardware acceleration.

Secure Retrieval-Augmented Generation on LLMs

May 2024 - Present

This project addresses the security vulnerabilities in Retrieval-Augmented Generation (RAG) systems for Large Language Models (LLMs).

Challenges

• RAG systems are vulnerable to poisoning attacks and jailbreaking attempts
• Traditional defense mechanisms often compromise model performance
• Need for robust security without sacrificing utility

Approach

• Implemented a novel objective function combining adversarial loss, BERTScore, and harmful loss
• Developed a robust RAG system for LLMs to counter universal attacks
• Created adaptive defense mechanisms that respond to attack patterns

Results

• Successfully prevented over 90% of state-of-the-art poisoning attacks
• Blocked jailbreaking attacks on RAG-based LLMs
• Maintained high performance on legitimate queries

TrojAI: Detecting Trojans in AI Models

2021 - 2023

Participated in the NIST TrojAI competition to detect backdoored or poisoned models among over 1000 adversarial training models.

Challenges

• Detecting subtle backdoors inserted during the training process
• Distinguishing between benign and malicious model behaviors
• Handling diverse model architectures and attack vectors

Approach

• Developed a two-stage detection flow: Polygon trigger detection and Instagram trigger detection
• Implemented advanced feature extraction techniques to identify suspicious patterns
• Created ensemble methods to improve detection accuracy

Results

• Achieved 2nd place out of 16 teams in Round 3 of the NIST TrojAI competition
• Exceeded target performance with cross-entropy loss < 0.3465
• Successfully identified various types of backdoor triggers

Learn more

Quantization of Deep Neural Networks

Feb 2021 - May 2021

Explored the application of Binary Neural Networks (BNNs) in oblivious inference to enable privacy-preserving machine learning.

Challenges

• Traditional privacy-preserving inference methods are computationally expensive
• Need for efficient cryptographic protocols for neural networks
• Balancing privacy, performance, and accuracy

Approach

• Enabled oblivious inference in Binary Neural Networks
• Explored the application of BNN in oblivious inference
• Devised lightweight cryptographic protocols tailored to BNNs

Results

• Achieved 2x faster inference compared to standard approaches
• Up to 11x faster inference for binary networks
• Maintained acceptable accuracy while providing privacy guarantees

View Paper

Optimization and Acceleration of Deep Learning

2021

Explored methods to optimize and accelerate deep neural networks for deployment on various hardware platforms, particularly resource-constrained devices.

Techniques Explored

• Pruning: Reduced model size by removing sections with little classification power
• Tucker Decomposition: Decomposed tensors into smaller core tensors and matrices
• Quantization: Reduced precision of weights and activations

Results

• Significant reduction in model size while maintaining accuracy
• Improved inference speed on resource-constrained devices
• Reduced power consumption for IoT and edge deployments

Learn more

Leakage Power Minimization

2020

Used Synopsys PrimeTime and PT-PX for timing and power analysis to implement strategies for gate sizing and Vt-swapping optimizations.

Approach

• Applied sensitivity function to balance leakage power and timing slacks
• Calculated sensitivity metrics based on ∆leakage_power / ∆slack
• Sized cells based on sensitivity ranking
• Used ECO flow with Cadence Innovus Implementation System

Results

Benchmark	Reduction Rate
usb_phy	82.92%
aes_cipher_top	73.68%
mpeg2_top	71.90%

Learn more

Ready to collaborate?

Let's discuss how my expertise in AI security and machine learning can benefit your organization.

Get in Touch Schedule a Call