Research Projects

Developed Beholder Signatures, a novel cryptographic signature framework designed to enhance transaction verification security and privacy in decentralized ecosystems. This project investigates mathematical bounds and lightweight verification structures for multi-party signatures.

Proposing a blockchain-agnostic, fully off-chain framework (UCRb) that enables flexible coin shifting (rebalancing) across multiple payment channels to solve liquidity management and channel depletion challenges in Payment Channel Networks.

Designing a privacy-preserving and reliable traffic information sharing framework (TrafficProof) for the Social Internet of Vehicles (SIoV). The system protects the privacy of vehicular telemetry and route data while ensuring auditability and validation of shared traffic event alerts.

The rapid growth of distributed renewable energy resources and smart grid technologies has accelerated the development of decentralized local energy markets, where prosumers and consumers can directly exchange electricity through peer-to-peer trading. While these markets offer improved energy efficiency, reduced transmission losses, and greater integration of renewable energy, they also introduce significant challenges related to data privacy, transaction security, scalability, and trust among participants. Privacy-preserving mechanisms are essential to protect sensitive user information, whereas scalable frameworks are required to efficiently support increasing numbers of users and transactions.

Privacy-preserving framework for secure verification in digital workflows. The goal is to design a system that can validate authenticity, authorization, and policy compliance while minimizing exposure of sensitive information. The work focuses on balancing security, privacy, and usability through a prototype implementation and performance evaluation. It also studies practical deployment considerations, including verification efficiency, data minimization, and scalability in real-world settings.

Investigated secure and privacy-preserving Retrieval-Augmented Generation (RAG) systems through extensive literature review and implementation analysis. Researched adversarial corpus poisoning attacks, statistical defense mechanisms, and cryptographic verification techniques, leading to the design of zk-CorpusGuard—a framework that combines Directional Relative Shift (DRS) and Zero-Knowledge Proofs (zk-SNARKs) to provide verifiable corpus admission, ensuring documents satisfy poisoning-detection policies before being indexed while preserving data privacy.

Developing intelligent techniques for detecting AI-generated synthetic voices, with a focus on enhancing the security, reliability, and trustworthiness of modern voice-based systems.

Developing a privacy-preserving and verifiable CRISPR off-target analysis framework using zero-knowledge proofs. The system enables patients to prove that exhaustive genome-wide off-target searches and scoring were performed correctly without revealing their genomic data. It ensures both computational integrity and genomic privacy, making secure genome editing workflows practical and trustworthy.

Employing federated learning in multi-UAV networks to facilitate decentralized intelligence through privacy-preserving collaborative training, secure model aggregation, robust communication, and protection against adversarial attacks without sharing sensitive local data.

Developing a semantic network traffic analysis framework for early-stage Advanced Persistent Threat (APT) detection using deep learning and semantic feature extraction. The framework identifies APT activities across the seven stages of the MITRE ATT&CK lifecycle to support proactive cyber defense.

Designing a robust, privacy-preserving Federated Learning (FL) architecture to enable collaborative machine learning across distributed nodes without exposing sensitive raw data. This project integrates Differential Privacy (DP), Secure Multi-Party Computation (SMPC), and Zero-Knowledge Proofs (ZKPs) to defend against gradient leakage, poisoning attacks, and membership inference threats. The framework optimizes communication efficiency and verification overhead to provide verifiable privacy guarantees for high-stakes decentralized training environments.

Developing a privacy-preserving federated learning framework for early-stage depression detection using passive behavioral sensing data (sleep, activity, phone usage, social interaction) from the StudentLife dataset. The framework combines clustered federated learning and Split Learning with local differential privacy, adaptive clipping, and DCT-based compression to protect sensitive mental health data while preserving model utility.

Investigating privacy-preserving WiFi sensing using Beamforming Feedback Information (BFI) for secure human activity recognition, with emphasis on attack analysis and defense mechanisms.

Investigating frameworks for federated unlearning to enable removing specific user data from trained machine learning models while maintaining the model's accuracy and ensuring rigorous privacy bounds.

Developing a neural differential equation solver using differentiable feature grids for accurate, scalable, and efficient learning of scientific simulation and PDE-based problems.

Developing an efficient framework for solving Partial Differential Equations (PDE) using gradient-augmented hash encoding to improve learning, accuracy, convergence speed, and computational efficiency.

Develop privacy-preserving deep learning frameworks for the early detection and classification of crop diseases using computer vision, edge-based mobile sensing, and distributed agricultural data. A Federated Learning approach enables farms, research centers, and mobile devices to collaboratively train a shared disease-detection model without transferring raw crop images to a central server. To protect local model updates during aggregation, Shamir Secret Sharing can divide each update into multiple cryptographic shares distributed among participating nodes. Only an authorized threshold of shares can reconstruct the aggregated result.

This project aims to address moderation architectures where moderators have absolute authority to issue state altering callbacks. Our work introduces a framework for federated threshold moderation to acheive cross platform accountability without sacrificing individual platform anonymity and user privacy. We propose a dual tier reputation architecture that reprimands malicious user activity. This project will actively research the tradeoffs of weighted trust thresholds to prevent malicious platform collusion, policy binding metadata to allow platforms to selectively filter global penalties and benchmark performance.

Investigating robust detection mechanisms for AI-generated deepfake images using advanced forensic analysis and deep learning techniques. This project focuses on developing generalizable classifiers capable of identifying subtle manipulation artifacts across diverse generative models, including GANs and diffusion-based architectures, to safeguard digital media integrity and counter misinformation threats.

zkViewAudit is a cryptographic framework for trustworthy digital media measurement that enables independent verification of audience metrics while preserving the privacy of viewers and publishers. Using Zero-Knowledge Proofs (ZKPs), it allows media platforms, advertisers, and auditors to verify impressions, views, and engagement without revealing sensitive user data or proprietary information. By ensuring bilateral accountability, zkViewAudit supports transparent, verifiable, and privacy-preserving media measurement.

Ethereum uses RANDAO, a distributed randomness beacon, to generate pseudorandom values for selecting future block proposers. Each proposer contributes a deterministic randao_reveal, which is mixed into the beacon state and influences subsequent validator assignments. In a RANDAO forking attack, a strategic validator does not merely withhold its own block. Instead, it selectively creates or supports an alternative chain that removes an honest proposer’s already-published block from the canonical chain. Because removing that block also removes its RANDAO contribution, the attacker obtains an alternative randomness outcome and may improve its probability of controlling profitable future proposer slots.

TriShieldRAG is a three-ring defense-in-depth framework designed to protect Retrieval-Augmented Generation (RAG) systems from knowledge corruption and adversarial corpus poisoning attacks. It deploys three independent defense rings across the pipeline: an Ingest Guard that screens documents for poisoning signatures, a Retrieval Scorer that re-ranks results based on provenance and consistency trust scores, and a Cross-LLM Consensus stage polling architecturally diverse LLMs (Claude, Mistral, Llama). TriShieldRAG significantly reduces attack success rates from ~91% down to ~13% while preserving accuracy on benign queries.