Discriminative adversarial networks for specific emitter identification

Abstract

The crucial issue in specific emitter identification (SEI) is the extraction of fingerprint features which can represent the differences among individual emitters of the same type. Considering that these emitters have the same intentional modulation on pulse, the fingerprint features originated from the unintentional modulation on pulse are extremely imperceptible and less detectable. However, existing feature extractions, either traditional handcrafted ones or deep learning based ones, have failed to ensure that their extracted features are rich in the unintentional modulation information (UMI) and not interfered by the intentional modulation information (IMI). To adequately take advantage of deep learning to address SEI, this Letter proposes a novel neural networks, named discriminative adversarial networks (DAN). By demarcating a clear boundary between IMI and UMI, DAN isolates IMI and thus reduces the burden of UMI mining during its feature extraction process. Experimental results demonstrate that DAN outperforms most methods in the literature.

A Novel Technique for Image Steganalysis Based on Separable Convolution and Adversarial Mechanism

“…Adversarial training is essentially a zero-sum game between the generator and the discriminator, which can usually be implemented as a complex min-max problem. We simplify this problem to a simple minimization problem by introducing GRL [22,26], which has no parameters associated with it. During the forward propagation, GRL acts as an identity transformation.…”

“…This paper draws on the ideas of transfer learning [22] to introduce adversarial training in steganographic detection, in order to suppress the content information of the image as much as possible and highlight the steganographic information to extract the steganographic embedding features more effectively. As shown in Figure 3, the adversarial mechanism can be divided into three parts: Label classifier: This part introduces GRL [22,26], which simplifies the adversarial training between the feature extractor and the classifier; it will optimize the loss function in the direction of negative gradient, and extract more image content features in order to mislead the discriminator's classification; thus, it is deemed more conducive to the detection of the existence of steganographic information. Through adversarial training, more steganographic embedding features can be isolated, and the accuracy of the network can be improved.…”

“…Label classifier: This part introduces GRL [22,26], which simplifies the adversarial training between the feature extractor and the classifier; it will optimize the loss function in the direction of negative gradient, and extract more image content features in order to mislead the discriminator's classification; thus, it is deemed more conducive to the detection of the existence of steganographic information. Through adversarial training, more steganographic embedding features can be isolated, and the accuracy of the network can be improved.…”

“…The training process of the network is as follows: an input image of 256 × 256 pixels, after passing through high-pass filters in the preprocessing layer, is sent to the subsequent separable convolution module and the basic convolution module for feature extraction; after a fully connected layer, the feature decomposition is used to split the feature into two parts, one of which needs to pass through the GRL [22,26], and is then sent to the subsequent fully connected layer for classification in order to complete the task of image steganalysis.…”

Research on emitter individual identification technology based on Automatic Dependent Surveillance–Broadcast signal

“…3 Talbot et al 4 designed a typical specific emitter identification/verification (SEI/SEV) processing prototype system, and a complete set of schemes for signal processing, parameter extraction, and identification library establishment is given, which provides a reference for signal processing flow of subsequent researchers. Chen et al 5 studied the interference of intentional modulation information (IMI) to unintentional modulation information (UMI) during the extraction of individual emitter features, and proposed discriminative adversarial networks (DANs), which determine the boundary between IMI and UMI; thus, in the process of individual feature extraction, the IMI interference is reduced. Prior research 6 established a signal model for the rising, stabilizing, and falling parts of the pulse waveform, converted the time-domain waveform of the signal into a two-dimensional binary image, and designed a new neural network to extract the subtle signal features.…”