The aim of the project is to implement a zkml framework on the fpga board and outperform the existing zkml frameworks running on cpu
All the following schemes were benchmarked on Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz. These were all tested on a standard configuration with no specialized hardware acceleration.
Here's a clean table format for:
| Scheme | Model | Prover Time (s) | Verifier Time (s) | Proof Size |
|---|---|---|---|---|
| vCNN(groth16) | vgg11 | 132.2 | 0.043 | 127b |
| zkCNN | Vgg11 | 20.4 | 2 | 109.906250 kb |
| MSCProof | vgg11 | 2.9 | 5 | 43kb |
| EZKL | lenet5 | 20.5 | 0.04 | 32b |
NOTE: zkCNN implementation is not zero knowledge hence the prover times are really not reflective with others
| Scheme | Prover Complexity | Verifier Complexity | Communication Complexity |
|---|---|---|---|
| vCNN | O(n log n) | O(log n) | O(1) |
| zkCNN | O(n^2) | O(1) | O(n^2) |
| MSCProof | O(n log^2 n) | O(log n) | O(nlogn) |
| EZKL | --- | O(log n) | O(1) |
The current plan is to implement the primitives of bn254 curve for a polynomial commitment scheme (PCS) and build up to a complete ZKML framework with two main plan of actions
PROS:
- Will significantly improve current framework
- Smaller proof size
CONS:
- Not flexible for all models
- CNNs are particularly challenging
PROS:
- Faster prover times
- Can handle more complex computations
CONS:
- Very large prover size
- High bandwidth requirements
find the msm synthesis report here