Large-scale deep neural networks (DNNs), such as large language models, have gained immense popularity due to their outstanding performance across various tasks. However, their application in resource-constrained scenarios faces significant challenges due to the high computational costs and memory usage of these models during inference. Model pruning emerges as a viable technique to mitigate these limitations by reducing the computational complexity of deep models. While existing research primarily focuses on maximizing inference efficiency without compromising accuracy, the privacy implications of pruning techniques remain largely unexplored. In this work, we systematically investigate the impact of popular pruning techniques on the privacy vulnerabilities of DNNs. We begin by applying common pruning schemes to various DNNs and evaluating their privacy risks, both before and after pruning, using model inversion attacks. We then analyze how the pruning rate and granularity affect these privacy vulnerabilities. Moreover, we conduct experiments on both original and pruned models equipped with defenses to confirm that the increase in privacy risks following pruning is not merely coincidental. Finally, we offer guidelines for the careful application of pruning techniques. Our findings serve as a cautionary note, highlighting the inherent privacy risks associated with current pruning schemes and providing valuable insights for developing pruning methods that are both efficient and secure.

This paper presents the design and implementation of a systematic Inter-Component Communications (ICCs) dynamic Analysis Technique (SIAT) for detecting privacy-sensitive data leak threats. SIAT’s specific approach involves the identification of malicious ICC patterns by actively tracing both data flows and implicit control flows within ICC processes during runtime. This is achieved by utilizing the taint tagging methodology, a technique utilized by TaintDroid. As a result, it can discover the malicious intent usage pattern and further resolve the coincidental malicious ICCs and bypass cases without incurring performance degradation. SIAT comprises two key modules: Monitor and Analyzer. The Monitor makes the first attempt to revise the taint tag approach named TaintDroid by developing the built-in intent service primitives to help Android capture the intent-related taint propagation at multi-level for malicious ICC detection. Specifically, we enable the Monitor to perform systemwide tracking of intent with five abstraction functionalities embedded in the interactive workflow of components. By analyzing the taint logs offered by the Monitor, the Analyzer can build the accurate and integrated ICC patterns adopted to identify the specific leak threat patterns with the identification algorithms and predefined rules. Meanwhile, we employ the patterns’ deflation technique to improve the efficiency of the Analyzer. We implement the SIAT with Android Open Source Project and evaluate its performance through extensive experiments on a particular dataset consisting of well-known datasets and real-world apps. The experimental results show that, compared to state-of-the-art approaches, the SIAT can achieve about 25% ∼200% accuracy improvements with 1.0 precision and 0.98 recall at negligible runtime overhead. Apart from that, the SIAT can identify two undisclosed cases of bypassing that prior technologies cannot detect and quite a few malicious ICC threats in real-world apps with lots of downloads on the Google Play market.

Recent years have seen increased attention to deep learning-based vulnerability detection frameworks that leverage neural networks to identify vulnerability patterns. Considerable efforts have been made; still, existing approaches are less ac-curate in practice. Prior works fail to comprehensively capture semantics from source code or adopt the appropriate design of neural networks. This paper presents SEVulDet, a Semantics-Enhanced learnable Vulnerability Detector that can accurately pinpoint vulnerability patterns by preserving path semantics into gadgets and learning from flexible-length codes. SEVulDet has two main characteristics: (i) SEVulDet employs a path-sensitive code slicing approach to extract sufficient path semantics and control flow logic into code gadgets. (ii) by inserting a spatial pyramidal pooling layer into the Convolutional Neural Network (CNN) with a well-designed multilayer attention mechanism, SEVulDet can handle gadgets of flexible-length semantics to avoid semantics loss incurred by traditional truncating or padding operations, and thus learn more potential vulnerability patterns. Comprehensive experimental results show that SEVulDet significantly outperforms classical static approaches and excels with state-of-the-art deep learning-based solutions by improving F1-measure to roughly 94.5%. Particularly, the elaborate design of the SEVulDet architecture helps us identify more real-world vulnerabilities than existing technologies.

An effective web attack detection method appears as a natural solution to protect web security, as they help to protect web applications. The traditional method of detecting web attacks is to encode the attack features manually into corresponding rules for detection. With the diversification of web attack methods, the demerits of the traditional methods have become increasingly noticeable. With the rapid development of high-performance computing and expansion of data volume, machine learning methods can obtain more efficient and accurate web attacks detection. In this paper, we exploit a bag of words based (BOW) model to extract features and further efficiently detect web attacks with hidden Markov algorithms. The experimental results show that, compared with the previous experiments of N-gram extraction feature algorithm, BOW has higher detection rate and lower false alarm rate with a lower cost. Finally, satisfactory results in the real environment are also achieved.