Poison in the Wires: Interactive Network Visualization of Data Attacks

Anya Vesna opens with a primer on machine learning and data poisoning, describing it as manipulation of training data that pushes a model toward unwanted behavior. She explains how poisoned data can enter through insider access to private datasets or through public sources that other models reuse. Chatbot incidents are used to show how quickly tampered input can shape output.

The middle section treats training data as a network of nodes and edges. The speaker presents GraphLeak, an open-source tool she built to be simpler to run in a browser. Users load node and edge files, choose which nodes or edges to poison, and compare clean and altered graphs side by side. Basic statistics and a rough risk rating accompany each comparison, with about ten percent poisoned points cited as the level where changes become noticeable. A live demo showed how deleting or adding nodes shifted the risk rating.

To close, the speaker argued that before-and-after comparisons, similar to a diff, can reveal tampering, and that provenance tracking helps. Suggested defenses include access control for private data and sanitization or rate limiting for public sources. Future work includes more statistical analysis and cycle detection. In the Q&A, she noted that judging whether new input is poisoned depends heavily on context, so guardrails around inputs matter.