Skills
Each area lists specific skills. Click a + to see the work where I used them.
-
Network protocols and encrypted traffic analysis
I read protocol behavior and application-layer patterns directly off the wire, encrypted or not, and work out what a network sensor can and cannot recover.
Where I used it
- ICS-Sniper (NDSS 2027). Profiled an industrial process from its encrypted traffic alone: extracted features from TLS- and VPN-encapsulated SCADA traffic, recovered control-loop periodicity with phase folding and superperiod detection, and clustered flows to find the packets that matter.
- Snoopy (IEEE TDSC 2022). Fingerprinted webpages from the sequence of encrypted resource sizes, for every user of a website at once, with about 90% accuracy across browsing contexts.
- Depending on HTTP/2 for Privacy? Good Luck! (DSN 2020). The first traffic-analysis attack to defeat HTTP/2 multiplexing.
- White Mirror (SIGCOMM 2019). Recovered the choices viewers make in interactive Netflix titles from encrypted traffic, 96% of the time in the worst case, and built the first interactive-video traffic dataset, from 100 viewers.
- AI-service fingerprinting (independent project, in progress). Classifying encrypted flows to AI services by provider, model and client surface, using paired HAR and pcap capture and a labelling set that covers more than 15 AI providers.
-
Attack research and proof-of-concept development
I take a weakness from root cause to working proof of concept, then measure its real impact and propose a countermeasure.
Where I used it
- ICS-Sniper (NDSS 2027). The first targeted blackhole attack on encrypted ICS traffic. From outside the plant perimeter, and without breaking encryption, it drops only the packets a process depends on, using NetfilterQueue on live traffic.
- Depending on HTTP/2 for Privacy? Good Luck! (DSN 2020). Built an active on-path adversary that adds jitter, throttles bandwidth and drops packets to force a server to send objects one at a time, which exposes their sizes.
- Security risks of AI/ML-enabled connected healthcare systems (CHASE 2024). Showed that adversarial glucose readings injected over a compromised Bluetooth link make a blood glucose management system's ML model mispredict. Success ranged from 44% to 100% depending on the patient.
- Snoopy and White Mirror. Side-channel attacks that leak user behavior from encrypted traffic without touching either endpoint.
-
Detection engineering and threat prevention
I turn an understanding of attacks into detection logic, and weigh the coverage and false-positive trade-offs that decide whether a detection ships.
Where I used it
- Learning from the Good Ones (DSML 2025). A risk-profiling defense that trains anomaly detectors on the data most resilient to attack. Against evasion attacks it raised recall by 27.5% for kNN and 16.8% for One-Class SVM.
- ICS-Sniper (NDSS 2027). Tested the attack against existing anomaly detectors, and evaluated network-layer mitigations such as traffic shaping for their coverage and overhead.
- Zeek and Kitsune network IDS with agentic triage (independent project, in progress). A live pipeline with Zeek for capture and features, Kitsune's autoencoder ensemble as the detector, and an LLM agent that turns per-packet anomalies into candidate Suricata rules and validates them.
- Detection characteristics framework (independent project, in progress). Specifies what each network detection observes, its evidence quality and its evasion surface, mapped to MITRE ATT&CK and ATLAS. Includes zeek-shadow-ai, a Zeek package for hunting unsanctioned AI traffic.
-
Security of AI and ML systems
I analyze how ML models fail under attack once they are wired into real devices and networks, and how to harden them.
Where I used it
- Security risks of AI/ML-enabled connected healthcare systems (CHASE 2024). Analyzed 20 FDA-approved ML-enabled medical devices, mapped known ML attacks to vulnerabilities in the peripherals they connect to, and demonstrated an end-to-end attack on a blood glucose management system.
- SAM (HealthSec 2024). A design-time method that anticipates false data injection attacks on ML-enabled medical devices, applied to two FDA-cleared devices.
- Systems-Theoretic and Data-Driven Security Analysis (Springer, invited). Extended that analysis with data from FDA and vulnerability databases; 11 of the 15 devices studied were susceptible to false data injection.
- Learning from the Good Ones (DSML 2025). Hardened DNN-based anomaly detectors against evasion attacks.
-
AI for security: LLM and agentic tooling
I build AI-assisted tools that take the manual effort out of security analysis, from vulnerability retrieval to alert triage.
Where I used it
- MedAIScout (Red Teaming GenAI workshop, NeurIPS 2024). A semi-automated pipeline, built on DistilBERT question answering and a locally hosted Llama 3, that reads a medical device's public documentation and retrieves the ML attack literature that applies to it. It found relevant vulnerabilities for four of five devices, in 10 to 20 minutes each.
- SAM (HealthSec 2024). Used GPT-4, GPT-4o and Llama 3 to generate attack steps from a system's design and its known vulnerabilities; between 84% and 96% of the generated steps were correct.
- Agentic alert triage (independent project, in progress). An LLM agent with a ReAct-style action space that works over anomaly-detector output, then writes and validates Suricata rules.
- Current work. Agentic, self-improving tools that identify vulnerabilities and attacks in networked multi-vendor systems.
-
Industrial control systems and OT security
I study how connecting industrial plants to the cloud changes their threat model, and test attacks and defenses on real controllers.
Where I used it
- ICS-Sniper (NDSS 2027). Led the five-member, multi-year project. Built hardware and software water-treatment testbeds with Allen-Bradley PLCs, and presented the work at USENIX Security '25 and, by invitation, at Idaho National Laboratory.
- Mitacs-SICI research internship (2021). Encrypted traffic analysis attacks on Internet-connected industrial control systems.
- Eradicator (2017). An integrated defense against cyber attacks on PLC-based industrial control systems. First prize at the Embedded Security Challenge, Cyber Security Awareness Week.
-
Threat modeling and risk assessment
I model threats for systems assembled from many vendors' parts, where the dangerous paths cross component boundaries.
Where I used it
- Security risks of AI/ML-enabled connected healthcare systems (CHASE 2024). Evaluated DREAD, STRIDE, FMEA and CVSS on ML-enabled connected medical systems, and showed that they miss risks that appear only when devices from several vendors are combined.
- SAM (HealthSec 2024). Brought STPA-Sec, a systems-theoretic security analysis, to ML-enabled medical devices at design time, and automated its most laborious steps.
- Systems-Theoretic and Data-Driven Security Analysis (Springer, invited). Combined that analysis with FDA device databases and CVE records to support security risk assessment before a device reaches the market.
- Postdoctoral work at UBC. Presented the findings to the FDA and to industry executives at Medtronic, GE and Medcrypt.
-
Privacy, IoT and cloud systems
I work on how personal data leaks, and how to keep it under control, in web, IoT and cloud applications.
Where I used it
- Turnstile (EuroSys 2026). Contributed the privacy research ideas for a hybrid information flow control framework for IoT applications. It found 190 privacy-sensitive data flows in 61 Node-RED applications, against 52 for CodeQL.
- ImmunoPlane (IoTDI 2024). Co-authored a middleware that keeps distributed IoT applications running through device failures, recovering within 2 seconds.
- Measuring quality of service in untrusted service chains (INFOCOM 2019). A scheme for verifying quality of service across service chains run by untrusted vendors, built on tamper-resistant evidence Bloom filters. It reached 97% prediction accuracy with about 7 times less storage than hash tables.
- ApproxBC (IEEE Communications Standards Magazine, 2018). Blockchain designs for devices with very little memory, using 8 times less storage than Merkle trees.
- Snoopy and White Mirror. Measured how much private behavior leaks from encrypted web and streaming traffic.
-
Programming and research engineering
I automate recurring analysis instead of repeating it by hand, and package research so that others can reproduce it.
Where I used it
- ICS-Sniper (NDSS 2027). Packet-capture pipelines in pyshark, tshark and Scapy, in-path live flow processing with NetfilterQueue, and an artifact package that covers testbed reproduction, network traces and the analysis algorithms.
- SAM and MedAIScout. Python tooling around hosted and locally run LLMs.
- AI-service fingerprinting (independent project, in progress). A capture harness built with httpx and Playwright that generates and labels real traffic.
- Zeek and Kitsune integration (independent project, in progress). A Python and Zeek integration layer for live interface capture, with Zeek cluster configuration on Ubuntu.
- SCION. Kernel implementation of a secure Internet communication protocol, a collaboration during my PhD.
- DSN 2024. Served on the artifact evaluation committee.
-
Research leadership, mentoring and communication
I take technical work from an open problem to a delivered result, and explain it to specialists, executives and regulators.
Where I used it
- Team leadership. Led a five-member research team for more than two years, owning scoping, the research roadmap, equipment procurement, testbed setup, experiment planning and publication.
- Funding. Co-authored a grant awarded CAD 500K by the National Cybersecurity Consortium of Canada, with quarterly reporting to the funder.
- Mentoring and teaching. Mentored 4 graduate students and 12 undergraduates and interns, and received three Star TA awards at IIT Madras.
- Peer review. Program committee member for ACM CCS 2024 to 2026 and ASIACCS 2026. ACM CCS 2024 Distinguished Reviewer, in the top 10% of 527 members.
- Industry and regulators. Worked directly with security practitioners at Medcrypt, Medtronic and GE, and with the FDA. Organized Birds-of-a-Feather sessions at USENIX Security 2023 and 2025.
- Speaking and writing. Invited talks at Idaho National Laboratory and ETS Montreal, a national science-writing award (AWSAR, top 100 of 4,993 entrants), and research covered by WIRED.