Balancing Thin Line Between Data Intelligence & Privacy
summary
Systematic overview claims Federated learning lets models train across isolated data while keeping raw data local and lists concrete attacks: inference attacks, model poisoning and data poisoning (label flipping and backdoor insertion). Defenses evaluated include differential privacy (adds noise and drops accuracy), cryptographic options like homomorphic encryption/SMC (keeps accuracy but raises compute), TEEs, and participant-level DP. Practical tools named: PyTorch-based frameworks, TensorFlow Federated and a Fate model stack; detection uses clustering/outlier rejection and error‑threshold rejects, but benchmarking and heterogeneous training remain open.
AI-written summary · about
chapters (9)
- 0:00Federated learning basics
- 1:08Federated learning constraints
- 6:02Data distribution models
- 11:15Privacy mechanisms
- 14:03Security challenges
- 20:25Training challenges
- 22:52Federated learning in finance
- 23:55Conclusion
- 24:16Q&A
related talks
09mInterview with Blaise Aguéra y Arcas
28mA Privacy Enhancing Federated Learning Framework for Multi-Party Distributed ..
27mCreate Value and Insight Through Privacy-preserving Data Collaboration [...]
34mDecentralized Machine Learning
41mFederated Learning: Machine Learning on Decentralized Data
15mLearning and Unlearning Your Data in Federated Settings
31mFrom Theory to Practice: Advances in Real-World Federated Learning
34mTraining without sharing: Is Federated Learning the answer to data privacy in AI?