Multi-label backdoor attack detection and identification method for privacy protection neural network model

Abstract

The invention discloses a multi-label backdoor attack detection and identification method for a privacy protection neural network model, and the method comprises the following steps: 1, carrying out the training of a privacy protection model through employing an SMPC technology in a three-party ciphertext environment, and generating an initialization function of a mask matrix Mask and a trigger matrix Pattern; 2, constructing an adversarial input TrojanInput function; 3, carrying out optimization training on a mask matrix and a trigger matrix, and optimizing a (mask, pattern) function; 4, constructing a trigger reverse engineering function; and 5, through a backdoor attack identification algorithm, carrying out an abnormal value detection function Backdoor Identificatio. The invention designs a backdoor attack detection scheme for detecting whether the model is attacked by the backdoor and which specific tag is attacked on the premise of protecting data and model parameter privacy.

Description

technical field [0001] The invention relates to the fields of cryptography, machine learning (deep learning) and machine learning security technology, in particular to the detection and identification of multi-label backdoor attacks on a neural network model with a privacy protection property that hides a backdoor. Background technique [0002] Deep neural networks (DNNs) have made amazing achievements in many fields, and applications based on DNNs are profoundly changing all aspects of our lives, such as medical diagnosis, autonomous driving, and image processing. The DNNs model can generally be obtained by training the training data set or retraining based on the pre-trained model. In order to obtain a more accurate model, it is usually necessary to obtain a large amount of data for model training. Due to the cost of time or money, it is difficult for users to complete DNN training on personal computers. Therefore, users want to outsource the model to cloud servers for t...

AI Technical Summary

Problems solved by technology

[0010] We can consider a two-party example, the participants Alice and Bob want to compare who is richer, they own their own property x and y, but they don’t want the other party to know their specific property

Images