Brief

Researchers have identified a critical vulnerability in Fully Homomorphic Encryption (FHE) schemes, specifically the widely used CKKS scheme, which can lead to overflow attacks. These attacks corrupt neural network outputs by causing inputs to exceed the tolerances of FHE circuits. To address this, the paper proposes a formal verification technique that calculates certified bounds for neuron ranges, effectively eliminating overflows and reducing failure rates to zero in experimental benchmarks. This overflow-free solution is compatible with existing CKKS frameworks by allowing the substitution of standard polynomials with rigorously designed ones. AI

Researchers have developed a new method for replacing the ReLU activation function in neural networks with quadratic polynomials, specifically for use with fully homomorphic encryption (FHE). This approach aims to reduce the computational cost of FHE-only inference by using lower-degree polynomials while preserving classification accuracy on calibration datasets. The method formulates the replacement as a linear separation problem and extends to cases with misclassified samples using convex hull relaxations, achieving faster inference times compared to existing methods. AI