Transpiler

Accepts standard ML model definitions written in frameworks like PyTorch, TensorFlow, or ONNX. The transpiler parses the computational graph, abstracting away the underlying cryptographic operations so developers do not need to write low-level FHE code.

• Converts torch.nn.Linear to encrypted matrix multiplication
• Handles control flow and tensor reshaping automatically
• Reduces the barrier to entry for privacy-preserving ML

Replaces non-linear activation functions like ReLU, Sigmoid, and GeLU with low-degree polynomial approximations. This is critical because FHE schemes natively support only addition and multiplication.

• Minimizes approximation error while respecting the noise budget
• Selects optimal polynomial degree for target precision
• Enables deep neural network evaluation on encrypted data

Injects bootstrap and relinearization operations at precise points in the circuit to prevent ciphertext corruption. The transpiler statically analyzes the multiplicative depth of the computation graph to ensure the noise level never exceeds the decryption threshold.

• Schedules bootstrapping to minimize latency overhead
• Applies modulus switching to extend computation depth
• Guarantees correct decryption of the final result

Leverages Single Instruction, Multiple Data (SIMD) encoding to pack thousands of plaintext values into a single ciphertext. The transpiler automatically tiles tensors and aligns data layouts to maximize parallel throughput.

• Amortizes the cost of homomorphic operations across data vectors
• Reduces ciphertext expansion and memory footprint
• Critical for achieving practical inference latency

Generates optimized circuits for specific FHE schemes like CKKS for approximate arithmetic or TFHE for fast boolean operations. The transpiler selects the most efficient scheme based on the computational workload and precision requirements.

• CKKS for high-throughput neural network inference
• TFHE for programmable bootstrapping and lookup tables
• BFV for exact integer computations

Ensures that the evaluated circuit reveals no information about the model architecture itself. The transpiler can inject noise flooding or circuit sanitization techniques to protect proprietary model weights during encrypted inference.

• Prevents model extraction attacks by the client
• Maintains IND-CPA security throughout evaluation
• Protects intellectual property in inference-as-a-service deployments