Search

CN-121984814-A - Robust safe transmission and identification method based on multi-teacher contrast countermeasure distillation

CN121984814ACN 121984814 ACN121984814 ACN 121984814ACN-121984814-A

Abstract

The invention provides a robust safe transmission and identification method based on multi-teacher contrast anti-distillation, and belongs to the technical field of automatic modulation and identification of radio signals. The method comprises the steps of enabling a sending end to perform key-driven disturbance filtering processing on a wireless signal to be transmitted based on a shared key to generate an encrypted signal and transmit the encrypted signal, enabling a receiving end to perform inverse disturbance reconstruction on the received encrypted signal based on the shared key to obtain a decrypted signal, constructing a multi-teacher distillation system comprising a plurality of teacher models and 1 student model, performing joint distillation training on the student model based on output and middle layer characteristics of the teacher models and the student model, performing joint distillation training comprising output distillation and characteristic comparison distillation, and dynamically weighting a plurality of loss items generated in the multi-teacher distillation process through an adaptive loss weight adjustment mechanism based on meta learning to obtain a modulation recognition result. The identification rate of illegal receivers is effectively reduced, and the performance of legal receiving ends under interference is improved.

Inventors

  • MA JITONG
  • DAI MINFENG
  • WANG JIE
  • Wan Jiachang
  • Jin Sinian
  • Ju Moran

Assignees

  • 大连海事大学

Dates

Publication Date
20260505
Application Date
20260128

Claims (10)

  1. 1. A robust security transmission and identification method based on multi-teacher contrast combat distillation, comprising: The transmitting end performs key-driven disturbance filtering processing on the wireless signals to be transmitted based on the shared key, generates encrypted signals and transmits the encrypted signals; The receiving end carries out inverse disturbance reconstruction on the received encrypted signal based on the shared secret key to obtain a decrypted signal; The multi-teacher distillation system is constructed and comprises a plurality of teacher models and 1 student model, wherein the combined distillation training is performed on the student model based on the output of the teacher models and the student models and the characteristics of the middle layer, and comprises the steps of outputting distillation and characteristic comparison distillation; and dynamically weighting a plurality of loss items generated in the multi-teacher distillation process through an adaptive loss weight adjustment mechanism based on meta-learning so as to obtain a modulation recognition result.
  2. 2. The method of claim 1, wherein the key driven perturbation filtering process comprises: generating complex frequency domain disturbance parameters by using a shared secret key as a randomness source through a cryptographically secure pseudo-random number generator, and mapping the disturbance parameters into amplitude and phase distribution of a frequency domain filter; and carrying out frequency domain disturbance on the original signal through the frequency domain filter.
  3. 3. The method of claim 2, wherein the key driven perturbation filtering process further comprises: and carrying out energy compensation on the disturbed signal to ensure that the encrypted signal and the original signal keep statistical consistency on the power spectral density distribution.
  4. 4. The method of claim 1, wherein the inverse disturbance reconstruction comprises: And updating the original spectrum structure of the restored signal through multiple rounds of frequency domain residual errors by adopting a frequency domain iterative reconstruction mode based on a shared key.
  5. 5. The method of claim 1, wherein the multi-teacher distillation system comprises: Training the first teacher model through the decrypted original sample to obtain high-discriminant feature representation; The second teacher model is trained on the decrypted challenge samples to obtain a high-robustness feature representation.
  6. 6. The method of claim 5, wherein the combined distillation training comprises: Based on probability distribution difference between teacher model output and student model output, knowledge distillation is carried out on the student model by using Kullback-Leibler divergence, and classification performance of the student model is constrained by combining supervision loss of a real label.
  7. 7. The method of claim 1, wherein the combined distillation training further comprises: based on the middle layer characteristics of the teacher model and the student model, a contrast learning mechanism is introduced, and characteristic alignment between the teacher and the students is realized by shortening the characteristic distances of similar samples and pushing away the characteristic distances of heterogeneous samples.
  8. 8. The method of claim 7, wherein the contrast learning mechanism adopts a generalized adaptive contrast learning manner to constrain the feature representations of the original sample and the challenge sample respectively by introducing a contrast mask related to the sample label.
  9. 9. The method of claim 1, wherein the meta-learning based adaptive loss weight adjustment mechanism comprises: And constructing a meta-weight prediction network, taking a plurality of loss values generated in the distillation process of a plurality of teachers as input, outputting corresponding loss weight parameters, and dynamically weighting training objective functions of the student model based on the loss weight parameters.
  10. 10. The method of claim 9, wherein the meta-weight prediction network is trained by meta-learning such that the loss weight parameters are adaptively updated as a function of training phase and data distribution.

Description

Robust safe transmission and identification method based on multi-teacher contrast countermeasure distillation Technical Field The invention relates to the technical field of automatic modulation and identification of radio signals, in particular to a robust safe transmission and identification method based on comparison of multiple teachers against distillation. Background The security of the communication link has been equally important as transmission efficiency and robustness in modern communication systems. Although the traditional data encryption method can guarantee communication safety to a certain extent, the strong encryption means are difficult to widely apply in the scene of limited resources due to the limitation of computational complexity. With the advent of new network architecture, some scenarios cannot directly employ traditional encryption techniques, and thus begin to explore utilizing the physical characteristics of wireless channels to enhance communication security. Physical layer security techniques form a cooperative guard mechanism with upper layer encryption by sending interfering signals to confuse potential eavesdroppers or reduce the decoding capability of unintended recipients. Meanwhile, the automatic modulation recognition technology is used as a key signal processing means in a complex electromagnetic environment, can recognize the modulation mode of a received signal in a multi-signal interference environment, and is widely applied to the fields of frequency spectrum management, interference recognition, signal monitoring and the like. The traditional AMR method relies on manual feature extraction, expert experience and priori knowledge are needed, the processing flow is complex, and the adaptability is limited. In recent years, deep learning models gradually become a mainstream scheme of modulation classification by virtue of their strong feature extraction capability and end-to-end architecture. For example, AMR model based on convolutional neural network verifies that high-precision recognition can be realized without artificial feature extraction, frame based on depth residual error network improves classification performance through signal matrix conversion and CNN combined training, and complex convolution and bidirectional LSTM model is combined to further improve recognition capability of complex modulation mode. However, depth models generally rely on gradient descent optimization parameters, and are susceptible to resistant attacks, resulting in significant degradation of the model under malicious interference. The existing defense means such as countermeasure training and countermeasure distillation can improve model robustness, but the problems of high calculation cost and low training efficiency are faced. Knowledge distillation is used as a light-weight solution to enhance generalization of student models by migrating teacher model knowledge, but knowledge expression of a single teacher model is still insufficient under complex channel conditions. Although multi-teacher knowledge distillation can integrate multi-source knowledge, the existing research has obvious blank in the field of modulation identification, and the multi-teacher method focuses on improving accuracy or single-field robustness, and is not applied to a physical layer encryption scene systematically. Furthermore, physical layer encryption techniques enhance security through key-driven computational complexity, but the depth model on which they rely may still be bypassed against attacks, resulting in new security threats to the system. Therefore, a robust security transmission and identification method based on multi-teacher contrast against distillation is needed. Disclosure of Invention In view of the above, the invention provides a robust safe transmission and identification method based on multi-teacher contrast anti-distillation, which aims at robust safe transmission of signals in complex electromagnetic environment, combines contrast learning technology on the basis of integrating physical layer encryption and multi-teacher distillation architecture, provides a multi-teacher contrast anti-distillation system for safe transmission of signals, and realizes characteristic alignment of teacher students by receiving contrast masks of decrypted signals, realizes defense against attacks, and combines the two to realize multiple defense in the signal transmission process. And constructing a cooperative defense system at the same time on the signal layer and the model layer, so as to realize good optimization of physical layer anti-eavesdropping and model layer anti-attack. For this purpose, the invention provides the following technical scheme: a robust security transmission and identification method based on multi-teacher contrast combat distillation, comprising: The transmitting end performs key-driven disturbance filtering processing on the wireless signals to be transmitted based on the sha