Poisoned Pickles - Security Risks and Protections for Serialized ML Models

Explore the security risks and protective measures associated with pickle serialization in machine learning during this 27-minute conference talk. Delve into the widespread use of the pickle module for serializing and distributing ML models, and understand the vulnerabilities that make it easy for attackers to inject arbitrary code into ML pipelines. Learn about the challenges in detecting poisoned pickles and discover emerging tools and techniques inspired by DevOps practices to generate safer, higher-quality pickles. Gain practical insights on how to protect your models from attacks and implement trust-or-discard processes to enhance the security of your ML workflows.