unknown

Federated learning and differential privacy promise to protect user data by training AI models without centralizing sensitive information, but research reveals significant vulnerabilities. Gradient inversion attacks can reconstruct training data from model updates, Byzantine attacks can poison shared models, and the privacy-utility tradeoff often disadvantages marginalized populations. Real-world deployments like Google's Gboard and the US Census Bureau's 2020 census demonstrate both benefits and limitations. Corporate incentives, regulatory gaps, and the complexity of mathematical guarantees create a disconnect between technical protections and user understanding. These techniques reduce certain risks while creating new attack surfaces, representing progress but not a complete solution to privacy in AI systems.

Federated Learning Under Fire: Why Your Data Still Leaks

Exploring the intersections of artificial intelligence, decentralised cognition, posthuman ethics, society and culture.