| （英） |
This talk presents a study on defense techniques against inversion attacks in split computing, where a pre-trained neural network is partitioned and executed across multiple devices. Inversion attacks attempt to reconstruct input data from intermediate layer outputs. Prior studies have mitigated such attacks by injecting noise into the intermediate representations, thereby balancing inference accuracy and attack resilience. However, these approaches typically estimate the required noise intensity based on diagonal Fisher information, which incurs significant computational overhead.
In this study, we propose a method that jointly performs lightweight estimation of the necessary noise intensity and generates a correction vector to compensate for accuracy degradation caused by noise injection. Experimental results demonstrate that the proposed method effectively reduces the accuracy of input reconstruction by adversaries while preserving the original inference performance, all with significantly lower computational cost compared to existing methods. |