Artificial intelligence (AI) server systems, including AI servers and AI server clusters, are widely utilized in AI applications. The performance of an AI server system determines the performance of the performance an AI application, which has garnered significant attentions and investments from industries and users. However, performance is influenced not only by the AI computing accelerating chips but also by the other configurations of an AI server system, such as architecture, memory, bus, central processing unit (CPU), interconnect equipment, software, etc. As these components are often provided by different vendors, the myriad combinations thereof present challenges in assessing performance in an architecture-neural, reproducible, fairness-enhanced, and performance bottleneck identification-oriented manner. In response to this need, this paper introduces AISBench, a performance benchmark for AI server systems. AISBench comprises standardized rules and a test toolkit that has been agreed upon by over 20 AI server system and server component manufacturers. Compared to other AI performance benchmarks, AISBench provides a more comprehensive metrics system that enables performance benchmarking as well as identification of performance bottlenecks for optimization. Experimental data indicate that our benchmark testing approach offers sufficient comprehensiveness, effectiveness, and stability.

Many edge computing applications based on computer vision have harnessed the power of deep learning. As an emerging deep learning model for vision, Vision Transformer models have recently achieved record-breaking performance in various vision tasks. But many recent studies on the robustness of the Vision Transformer have shown that the Vision Transformer is still vulnerable to adversarial attacks and is easily affected by adversarial attacks, causing the model to misclassify the input. In this work, we ask an intriguing question: “Can Adversarial Perturbations against Vision Transformers be detected with model explanations?” Driven by this question, we observe that benign samples and adversarial examples have different attribution maps after applying the Grad-CAM interpretability method on the Vision Transformer model. We demonstrate that an adversarial example is a Feature Shift of the input data, which leads to an Attention Deviation of the visual model. We propose a framework for capturing the Attention Deviation of vision models to defend against adversarial attacks. Furthermore, experiments show that our model achieves expectative results.

Context-based paring is a promising direction for pairing IoT devices constrained in user interfaces (UIs). However, it takes a proximate distance or a long time for IoT devices to sense highly correlated context with enough entropy. In this work, we present a fast and secure approach, namedMPairing, to pairing multiple commercial off-the-shelf (COTS) IoT devices. This approach is based on the key idea that devices co-located within aphysically-secure boundarycan perceive qualified context under the help of human-in-the-loop (HITL). Specifically, we leverage received-signal-strength (RSS) trajectory data with manually-generated interference in a short period as the shared secret to achieve fast and secure pairing. Subsequently, the real-time RSS trajectory data is utilized to generate random numbers in lieu of pre-shared key (PSK), which makes our scheme more resistant to background attacks. We theoretically prove the security of our pairing scheme and implement it in real-world environments. Our experimental results demonstrate that our scheme can effectively defend against malicious devices by imposing a threshold on the similarity of RSS trajectory data. The experimental results also show that, compared with the traditional context-based pairing that takes up to 24 hours, in our scheme it takes only 10 seconds on average for a legitimate device to pass the similarity checking, which is efficient and robust.

Since big data becomes a main impetus to the next generation of IT industry, data privacy has received considerable attention in recent years. To deal with the privacy challenges, differential privacy has been widely discussed and related private mechanisms are proposed as privacy-enhancing techniques. However, with today’s differential privacy techniques, it is difficult to generate a sanitized dataset that can suit every machine learning task. In order to adapt to various tasks and budgets, different kinds of privacy mechanisms have to be implemented, which inevitably incur enormous costs for computation and interaction. To this end, in this article, we propose two novel schemes for outsourcing differential privacy. The first scheme efficiently achieves outsourcing differential privacy by using our preprocessing method and secure building blocks. To support the queries from multiple evaluators, we give the second scheme that employs a trusted execution environment to aggregately implement privacy mechanisms on multiple queries. During data publishing, our proposed schemes allow providers to go off-line after uploading their datasets, so that they achieve a low communication cost which is one of the critical requirements for a practical system. Finally, we report an experimental evaluation on UCI datasets, which confirms the effectiveness of our schemes.

Existing system dealing with online complaint provides a final decision without explanations. We propose to analyse the complaint text of internet fraud in a fine-grained manner. Considering the complaint text includes multiple clauses with various functions, we propose to identify the role of each clause and classify them into different types of fraud element. We construct a large labeled dataset originated from a real finance service platform. We build an element identification model on top of BERT and propose additional two modules to utilize the context of complaint text for better element label classification, namely, global context encoder and label refiner. Experimental results show the effectiveness of our model.

The alarming growth rate of malicious apps has become a serious issue that sets back the prosperous mobile ecosystem. A recent report indicates that a new malicious app for Android is introduced every 10 s. To combat this serious malware campaign, we need a scalable malware detection approach that can effectively and efficiently identify malware apps. Numerous malware detection tools have been developed, including system-level and network-level approaches. However, scaling the detection for a large bundle of apps remains a challenging task. In this paper, we introduce Significant Permission IDentification (SigPID), a malware detection system based on permission usage analysis to cope with the rapid increase in the number of Android malware. Instead of extracting and analyzing all Android permissions, we develop three levels of pruning by mining the permission data to identify the most significant permissions that can be effective in distinguishing between benign and malicious apps. SigPID then utilizes machine-learning-based classification methods to classify different families of malware and benign apps. Our evaluation finds that only 22 permissions are significant. We then compare the performance of our approach, using only 22 permissions, against a baseline approach that analyzes all permissions. The results indicate that when a support vector machine is used as the classifier, we can achieve over 90% of precision, recall, accuracy, and F-measure, which are about the same as those produced by the baseline approach while incurring the analysis times that are 4-32 times less than those of using all permissions. Compared against other state-of-the-art approaches, SigPID is more effective by detecting 93.62% of malware in the dataset and 91.4% unknown/new malware samples.