Recsperts - Recommender Systems Experts #4: Adversarial Machine Learning for Recommenders with Felice Merra
Feb 23, 2022
Felice Merra, an applied scientist at Amazon, discusses Adversarial Machine Learning in Recommender Systems. Topics include perturbing data and model parameters, defense strategies, motivations for attacks, and privacy-preserving learning. The goal is to make systems more robust against potential attacks. They also touch on the challenges of robustifying multimedia recommender systems.
AI Snips
Chapters
Transcript
Episode notes
Catalog Size and Attack Difficulty
- Attacking recommender models aims to profit by promoting specific items.
- Larger catalogs make attacks harder; smaller catalogs are easier targets.
Adversarial Images
- Adversaries can upload adversarial images, like a gun disguised as a toy.
- This manipulates recommendations, potentially showing guns to children.
Adversarial Machine Learning in Recommender Systems
- Adversarial machine learning investigates vulnerabilities in recommender systems.
- Attackers manipulate interaction data, metadata, or model parameters.