Andrei Stoian - Open-source Machine Learning on Encrypted Data | PyData Amsterdam 2024

Fully Homomorphic Encryption (FHE) enables processing encrypted data without decryption, providing security while data remains usable

ConcreteML is an open-source library that makes machine learning on encrypted data accessible, mimicking familiar frameworks like PyTorch and scikit-learn

Key FHE operations include:

• Addition of encrypted values
• Table lookups
• Conversion of floating-point operations to integer operations
• Working with quantized values (typically 8-16 bits)

Performance considerations:

• 1,000-10,000x computation overhead compared to unencrypted operations
• Ciphertext size can be up to 1000x larger than cleartext
• Compression can reduce expansion factor to 10-20x
• Latency of a few seconds is typical

Applications include:

• Private inference
• Spam filtering
• DNA ancestry analysis
• LLM fine-tuning
• Secure data marketplaces
• Healthcare data processing

Machine learning models supported:

• Linear models
• Decision trees
• Neural networks
• Large Language Models (with distributed computation)

Implementation requires:

• Representative training data for calibration
• Quantization parameters optimization
• Noise management in encrypted computations
• Converting floating point to integer operations

Security benefits:

• Data remains encrypted during processing
• Only the key holder can decrypt results
• Protection against data leaks
• Reduced regulatory compliance burden