Doctoral Thesis Proposal - Saranya Vijayakumar

— 1:30pm

Location:
In Person - Reddy Conference Room, Gates Hillman 4405

Speaker:
SARANYA VIJAYAKUMAR , Ph.D. Student
Computer Science Department
Carnegie Mellon University
https://svijayakumar2.github.io/index.html

Protection Boundary Integrity: Detecting and Preventing Security Failures Across Contexts

Modern computational systems deploy technical guardrails to enforce security, privacy, and safety boundaries across increasingly complex operational contexts. While effective within their design contexts, these mechanisms exhibit systematic vulnerabilities when systems transition between different operational modes: across interaction modalities, through temporal evolution, or when integrating neural and symbolic reasoning. This dissertation investigates where, how, and why security mechanisms fail at these critical transitions.

First, I demonstrate patterns of boundary failure through empirical analysis across multiple domains. My cross-modal work evaluates such failures in browser-agent safety auditing (BrowserART) and authenticity detection of AI-generated code (CodeFusion). Through BrowserART, I show that language models refusing harmful instructions in chat interfaces pursue identical harmful behaviors when deployed as browser agents, despite identical safety training. Through CodeFusion, I analyze visual structure and semantic content, demonstrating that authenticity boundaries require reasoning across representational modalities. Second, I identify temporal vulnerabilities that emerge when security mechanisms designed for static analysis cannot adapt to evolving threats. I demonstrate this through MalCentroid, tracking malware family evolution while maintaining robustness against adversarial obfuscation, and through graph-based fraud detection systems identifying attack patterns emerging across temporal transaction sequences. Through TRACE, I achieve successful re-identification against Google's Topics API by exploiting vulnerabilities where privacy mechanisms protecting individual observations fail when adversaries analyze sequential behavioral patterns.

Finally, I introduce methods to bridge neural-symbolic security boundaries. Through SMTLayer, I integrate satisfiability solvers directly into neural architectures, achieving substantial data efficiency improvements while maintaining formal logical guarantees. In my proposed work, I introduce verifiable protection mechanisms for language models through a game-theoretic prover-verifier framework and develop multiplicative gating architectures enabling efficient learning of complex logical structures like XOR gates that standard architectures struggle to represent. This research provides foundations for building protection mechanisms that maintain integrity across the complex operational transitions required for safe deployment of autonomous computational systems.

Thesis Committee
Christos Faloutsos (Co-Chair)
Matt Fredrikson (Co-Chair)
Sarah Cen
Mihai Christodorescu (Google Research)

Additional Information 
 

For More Information:
matthewstewart@cmu.edu

Add event to Google
Add event to iCal