Adversarial Memory Research Articles

An undercarriage image driven anomaly detection method for metro vehicle based on adversarial memory enhancement

Anomaly detection is essential to ensure metro vehicles' safe operation. Error reconstruction-based anomaly detection methods have been widely studied because they only need to be trained by normal data and do not require much anomaly data, which is challenging to obtain. However, sometimes the auto-encoder network for error reconstructing “generalizes” so well that it also rebuilds the anomaly well, leading to missed anomaly detection. Therefore, this paper proposes an undercarriage image-driven anomaly detection method for metro vehicles based on adversarial memory enhancement. Firstly, this study performs component segmentation based on YOLOv5 detection results and constructs a component anomaly detection dataset. Secondly, an anomaly detection method based on memory enhancement and adversarial training of encoding-decoding-encoding structure is proposed for component anomaly detection. It enables the auto-encoder to reconstruct the image better. Thirdly, the combined indicator of the difference between potential features and reconstruction error is used as an anomaly indicator for anomaly detection of metro components, reducing the rate of fault misses. The experimental results on the established dataset demonstrate that the proposed method reduces false negative rates of 92.4%, 92.6%, 74.6%, and 59.1% compared with [Formula: see text], [Formula: see text], GANomaly, and MemAE, respectively.

Unsupervised Anomaly Detection Approach Based on Adversarial Memory Autoencoders for Multivariate Time Series

The widespread usage of Cyber Physical Systems (CPSs) generates a vast volume of time series data, and precisely determining anomalies in the data is critical for practical production. Autoencoder is the mainstream method for time series anomaly detection, and the anomaly is judged by reconstruction error. However, due to the strong generalization ability of neural networks, some abnormal samples close to normal samples may be judged as normal, which fails to detect the abnormality. In addition, the dataset rarely provides sufficient anomaly labels. This research proposes an unsupervised anomaly detection approach based on adversarial memory autoencoders for multivariate time series to solve the above problem. Firstly, an encoder encodes the input data into low-dimensional space to acquire a feature vector. Then, a memory module is used to learn the feature vector’s prototype patterns and update the feature vectors. The updating process allows partial forgetting of information to prevent model overgeneralization. After that, two decoders reconstruct the input data. Finally, this research uses the Peak Over Threshold (POT) method to calculate the threshold to determine anomalous samples from normal samples. This research uses a two-stage adversarial training strategy during model training to enlarge the gap between the reconstruction error of normal and abnormal samples. The proposed method achieves significant anomaly detection results on synthetic and real datasets from power systems, water treatment plants, and computer clusters. The F1 score reached an average of 0.9196 on the five datasets, which is 0.0769 higher than the best baseline method.

Defending against adversarial attacks using spherical sampling-based variational auto-encoder

Although deep neural networks achieve outstanding performance in many tasks, adding very imperceptible perturbations to clean images can easily fool the deep neural network. In this paper, we propose a new defence model: Adversarial Memory Variational AutoEncoder(AdMVAE), that can be used to transform adversarial images into clean images. At inference time, it finds an output that is similar to a given image in a high probability region of the manifold space. And the memory module uses normal features to reconstruct the image in the process of reconstruction. It can effectively prevent the reconstruction of malicious perturbations and avoid defense failure. Our approach is a pre-processing module that does not change the results of the classifier. Therefore, it can be combined with other defence models to jointly improve the performance robustness of the classifier. The experimental results on three benchmark datasets including Fashion-MNIST, CIFAR10 and Imagenet show that the proposed method outperforms the state-of-the-art defense methods.

Adversarial memory for detecting destructive races

Multithreaded programs are notoriously prone to race conditions, a problem exacerbated by the widespread adoption of multi-core processors with complex memory models and cache coherence protocols. Much prior work has focused on static and dynamic analyses for race detection, but these algorithms typically are unable to distinguish destructive races that cause erroneous behavior from benign races that do not. Performing this classification manually is difficult, time consuming, and error prone. This paper presents a new dynamic analysis technique that uses adversarial memory to classify race conditions as destructive or benign on systems with relaxed memory models. Unlike a typical language implementation, which may only infrequently exhibit non-sequentially consistent behavior, our adversarial memory implementation exploits the full freedom of the memory model to return older, unexpected, or stale values for memory reads whenever possible, in an attempt to crash the target program (that is, to force the program to behave erroneously). A crashing execution provides concrete evidence of a destructive bug, and this bug can be strongly correlated with a specific race condition in the target program. Experimental results with our Jumble prototype for Java demonstrate that adversarial memory is highly effective at identifying destructive race conditions, and in distinguishing them from race conditions that are real but benign. Adversarial memory can also reveal destructive races that would not be detected by traditional testing (even after thousands of runs) or by model checkers that assume sequential consistency.

Universal fault tolerant quantum computation on bilinear nearest neighbor arrays

Assuming an array that consists of two parallel lines of qubits and that permits only nearest neighbor interactions, we construct physical and logical circuitry to enable universal fault tolerant quantum computation under the $[[7,1,3]]$ quantum code. A rigorous lower bound to the fault tolerant threshold for this array is determined in a number of physical settings. Adversarial memory errors, two-qubit gate errors and readout errors are included in our analysis. In the setting where the physical memory failure rate is equal to one-tenth of the physical gate error rate, the physical readout error rate is equal to the physical gate error rate, and the duration of physical readout is ten times the duration of a physical gate, we obtain a lower bound to the asymptotic threshold of $1.96\times10^{-6}$.

On the Power of Quantum Memory

We address the question whether quantum memory is more powerful than classical memory. In particular, we consider a setting where information about a random n-bit string X is stored in s classical or quantum bits, for s<n, i.e., the stored information is bound to be only partial. Later, a randomly chosen predicate F about X has to be guessed using only the stored information. The maximum probability of correctly guessing F(X) is then compared for the cases where the storage device is classical or quantum mechanical, respectively. We show that, despite the fact that the measurement of quantum bits can depend arbitrarily on the predicate F, the quantum advantage is negligible already for small values of the difference n-s. Our setting generalizes the setting of Ambainis et al. who considered the problem of guessing an arbitrary bit (i.e., one of the n bits) of X. An implication for cryptography is that privacy amplification by universal hashing remains essentially equally secure when the adversary's memory is allowed to be quantum rather than only classical. Since privacy amplification is a main ingredient of many quantum key distribution (QKD) protocols, our result can be used to prove the security of QKD in a generic way.