Disturbance detection device for aviation training data subjected to poisoning attack

The invention provides a device for detecting disturbance of aviation training data subjected to poisoning attack. Comprising the steps of S1, data acquisition; s2, constructing a neural network: zooming the input image to a uniform size, and then performing feature extraction through two branches on the left side and the right side to obtain a new feature map fused with fractal geometric features; s3, an attack process; s4, model training; s5, model evaluation and optimization; and S6, judging whether the engine has defects or not on the new picture by using the model. Through the above steps, a real attack scene can be simulated, and then the performance of the model in the face of malicious tampering of data is mined and analyzed by designing an iterative algorithm for multiple times of adjustment and optimization, and finally the security and robustness evaluation of the model is obtained. The problem that a comprehensive and practical evaluation means is lacked in the safety evaluation field in an aviati