Projects

Applied AI / ML projects.

Projects spanning hardware security, adversarial ML, LLM-assisted explainability, hyperdimensional computing, drug discovery, and ML systems. Each entry shows the problem, approach, result, and stack.

Featured Deep Dives

Selected research project deep dives.

Expanded context for selected work in hardware Trojan detection, adversarial attacks on ML-based security models, and scientific machine learning.

Golden-free hardware Trojan detection and localization workflow visual — Open full figure

Hardware Security Golden-Free Detection

Golden-Free AI-Assisted Hardware Trojan Detection

Hardware Trojans are malicious IC modifications that can leak sensitive information, disrupt critical systems, or damage trust in the semiconductor supply chain. Many side-channel detection methods depend on golden reference chips, which are expensive, unavailable, or impractical in outsourced fabrication settings.

My work uses Ring Oscillator Network side-channel measurements with unsupervised clustering to detect Trojan-affected chips without golden data. The published work demonstrated strong detection behavior for small, short-triggered Trojans, reaching 93% accuracy while keeping the method practical for supply-chain validation.

A related submitted extension combines golden-free clustering with on-chip localization. Instead of only saying that a chip is suspicious, the framework estimates the likely Trojan-affected region so reverse engineering can focus on a smaller part of the die. FPGA validation on MD5 processor prototypes reached 95% accuracy across multiple Trojan sizes and designs.

93% published accuracy 95% submitted extension Golden-free SCA

Read ACM JETC paper

Drug Discovery IJMS, 2022

CSatDTA: Drug-Target Affinity Prediction

Drug discovery brings together chemistry, pharmacology, and biology to identify potential treatments. In early-stage drug development, predicting drug-target affinity is important because it helps estimate how strongly a drug candidate may interact with a biological target.

CSatDTA combines convolution with self-attention for molecular drug and target sequences. Convolution captures local sequence patterns, while self-attention helps model long-range interactions that standard convolution-only approaches can miss.

Comparative experiments showed that CSatDTA outperformed previous sequence-based and related approaches, demonstrating strong retention ability for drug-target affinity prediction.

Read IJMS paper

Adversarial attack flow against clustering-based golden reference-free hardware Trojan detection — Open full figure

Adversarial ML IEEE, 2024

Adversarial Attack Against Golden Reference-Free Hardware Trojan Detection Approach

Golden reference-free hardware Trojan detection methods use unsupervised learning to identify suspicious ICs from untrusted manufacturing flows without relying on clean reference chips. This project asks what happens when the clustering model itself becomes the target of an adversarial attack.

The work introduces feature-space adversarial attacks against K-means clustering models trained on frequency side-channel analysis data. By generating adversarial samples, an attacker can alter Trojan behavior so that the clustering model assigns the source sample to a targeted cluster.

The study showed that the clustering model is highly vulnerable, with a 99% success rate in misleading the detector. The result is important because it exposes a weakness in otherwise practical golden reference-free detection pipelines and motivates robustness-aware hardware-security ML.

99% attack success rateK-means clusteringFrequency SCA

Read IEEE paper

More Projects

Additional research and applied ML work.

AI + Signal/Image Processing Integration, 2026

AI-Enabled Image Processing for Hardware Trojan Identification

Challenge: Stealthy IC modifications threaten semiconductor supply-chain trust, while functional testing and reverse engineering are costly to scale.

Build: Converted side-channel data into image-like representations and used image processing with unsupervised machine learning to cluster Trojan behavior.

Impact: The publisher page reports 95% hardware Trojan detection accuracy using real hardware side-channel data.

Pythonscikit-learnDSPImage ProcessingFPGA

ScienceDirect

Hardware Security / Adversarial ML Microelectronics, 2026

Adversarial Attack Resilient ML-Assisted Golden Free Hardware Trojan Detection

Challenge: Security models that perform well under clean conditions can fail under perturbations, process noise, or adaptive attacks.

Build: Used statistical feature extraction, clustering, internal filtering, and pseudo-labeling to train a detector from reliable chip patterns.

Impact: Experimental evaluation showed high detection accuracy with zero false negatives and resilience against adversarial perturbations.

PythonARTscikit-learnAdversarial MLClusteringpandas

Microelectronics DOI

Journal of Electronic Testing / Under Submission Manuscript under submission

Synthetic Data Augmentation for Robust Hardware Trojan Detection

Challenge: Hardware Trojan datasets are often small, imbalanced, and vulnerable to adversarial perturbations, making robust evaluation difficult.

Build: Combined feature-space attack analysis with synthetic data augmentation and adversarially trained models for side-channel Trojan detection.

Impact: Developed into a Journal of Electronic Testing submission that extends the hardware-security research story beyond clean-data detection.

PythonSMOTETVAESVMAdversarial TrainingSide-Channel Data

Healthcare AI / LLMs Active AI prototype

Clinical Communication and LLM-Based HEART Rubric Scoring

Challenge: Medical error disclosure training needs transparent, clinically meaningful assessment of resident-patient conversations.

Build: Developed an AI-compatible HEART rubric schema with transcript-only 0-3 scoring, evidence spans, and validation planning against human ratings.

Impact: Positioned as an interdisciplinary research MVP and collaboration-ready prototype, not a deployed clinical product.

PythonPyTorchTransformersLongformerSBERTJSON Schemas

Contact

HDC / Cybersecurity IEEE Access / SATC

Hyperdimensional Computing for IoT Anomaly Detection

Challenge: IoT intrusion detection needs efficient processing of high-dimensional network data while detecting known and unknown attack patterns.

Build: Applied hyperdimensional representations to network anomaly detection, including NSL-KDD and hybrid HDC-based frameworks.

Impact: Related public records describe HDC-based IoT anomaly detection and D2H-AD as a hybrid model for advanced anomaly detection.

PythonHDCML EvaluationNSL-KDDAnomaly Detection

arXiv IEEE DOI

LLMs / XAI / IoT Security SATC, 2025

LLM-Assisted Explainable IoT Systems for Critical Infrastructure

Challenge: Security monitoring often generates anomaly scores and alerts that are difficult for human analysts to interpret.

Build: Combined numerical anomaly detection with LLM-assisted preprocessing and explanation generation for human interpretability.

Impact: Conference work explores LLM-assisted explainability for cybersecurity in critical infrastructure settings.

PythonLLM PromptingAutoencodersXAIIoT Datasets

IEEE DOI

AI for Software Security IEEE Access, 2025

AI for Reverse Engineering and Software Security Survey

Challenge: Reverse engineering and software analysis are labor-intensive, yet increasingly supported by AI systems.

Build: Surveyed AI applications in decompilation, malware analysis, binary analysis, vulnerability discovery, and security automation.

Impact: Published in IEEE Access with a taxonomy of AI-assisted reverse engineering and software security research.

Literature ReviewSecurity AnalysisReverse EngineeringAI Systems

IEEE DOI

Distributed AI / Privacy DCAS, 2023

Secure and Privacy-Aware AI Hardware with Federated Learning

Challenge: Security-aware AI hardware must balance privacy, distributed learning, and practical implementation constraints.

Build: Implemented secure and privacy-aware AI hardware workflows using distributed federated learning concepts.

Impact: Conference publication and graduate research direction in distributed privacy-aware AI systems.

PythonFederated LearningDistributed SystemsAI Hardware

Scholar

Systems / ML Performance MLKE, 2024

Parallel Supervised ML in Multicore Environments

Challenge: Large ML experiments need reproducible, faster execution across multicore environments.

Build: Designed a parallel approach to improve supervised ML performance in a multicore setting.

Impact: Published work supports Ashutosh's builder credibility across ML experimentation and systems optimization.

PythonParallel ComputingML PipelinesPerformance Evaluation

Scholar