中央研究院 資訊科學研究所

Support Vector Machine (SVM) is a promising machine learning method, benefiting from its famous kernel trick. However, recent research pointed out the security concern for kernels in distributed environments. The private data hided in kernel matrix may be disclosed and leaked by the insider collusion attack. For countering this attack, we propose a hybrid solution based on matrix-gaussian differential privacy and homomorphic proxy re-encryption. Our experimental results show that not only is data privacy preserved by differential privacy but the accuracy result is also made lossless and efficient by homomorphic proxy re-encryption encryption.