Interested in what breaks — and why.
I'm Harsh — a third-year Computer Science & Business Analytics student at FLAME University, Pune. I have a deep interest in cybersecurity: not just in the tools, but in understanding how systems fail, where trust gets misplaced, and what it actually takes to build something genuinely secure.
Currently interning at C3iHub, IIT Kanpur, working on cybersecurity for the Maharashtra Metro project in Nagpur.
Discovered two security vulnerabilities in FLAME University's Moodle-based learning management system during independent security testing. The first was a brute-force login vulnerability caused by absent rate limiting; the second, a session hijacking risk through plaintext cookies exposed over unencrypted connections. Both were documented with reproducible proof-of-concept write-ups and reported through responsible disclosure.
Built and configured a signature-based intrusion detection system using Snort on a Kali Linux virtual machine. Wrote custom detection rules targeting common attack patterns including port scans, brute-force attempts, and suspicious payloads, with real-time alerting for flagged traffic.
Developed a two-part security tool: a login system with OTP-based multi-factor authentication delivered through AWS SES, featuring anti-brute-force lockout and DynamoDB credential storage — and a companion password manager using AES encryption for local vault security.
Simulated the Open Shortest Path First routing protocol using Dijkstra's algorithm in a custom network topology. The implementation models neighbor discovery, link-state advertisement exchange, and shortest-path routing table computation.
Most deepfake detection models focus on a binary question: is this face real or fake? What they don't account for is how the subject's facial expression — happiness, anger, surprise — affects that judgment. My research investigates emotion-wise bias in deepfake detection, a genuinely underexplored problem with real implications for the fairness of automated media forensics.
Working under Prof. Manoranjan Dash (Dean, School of Computing & Data Sciences, FLAME University), I led a four-member team through the full research pipeline: data acquisition, preprocessing, model evaluation, and paper drafting. Using frequency-domain modeling and a fine-tuned ResNet CNN, we improved baseline detection accuracy from 50% to 75% and reduced emotion-specific bias by 9% — all while maintaining strong F1-scores across emotional categories.
The work is ongoing. The goal is a detection framework that doesn't just catch deepfakes, but catches them consistently regardless of what the face in the video is doing.
What I'm up to these days, inspired by nownownow.com.
Last updated May 2026
I write occasionally about security, systems, and things I find interesting. Coming soon.