Forest Global Training Partners

Cyber Risks · Unit 42 Palo Alto Networks

The Innocent Until Proven Guilty Learning Framework Helps Overcome Benign Append Attacks

Summary: Unit 42 researchers presented a prototype-based learning framework called Innocent Until Proven Guilty (IUPG) at the 4th Deep Learning and Security Workshop, co-located with the 42nd IEEE Symposium on Security and Privacy, aimed at making neural network classifiers robust to benign content injected into inputs. Their JavaScript malware classifiers built with categorical cross-entropy loss achieved well over 99% accuracy on a test set, yet 10,000 characters of random benign content appended to malicious samples flipped the verdict more than 50% of the time.

Key Takeaway: An attacker needs no knowledge of the victim classifier and only plentiful, trivially produced benign content to defeat a model that is otherwise over 99% accurate, and Palo Alto Networks reports detecting over 130,000 malicious items from mid-April to mid-May using IUPG-trained models.

Read the original at Unit 42 Palo Alto Networks ↗

Topics
Malware
Related