Decompilation plays a critical role in firmware analysis and reverse engineering by enabling the recovery of high-level source code from binary executables. However, existing neural decompilation models often face challenges due to the semantic gap between assembly code and high-level languages, and they typically require large-scale models that impose significant computational demands. In this paper, we propose DECodeT5, a lightweight and efficient neural decompilation method for the C language. DECodeT5 builds upon the CodeT5 code generation model by integrating a pre-trained assembly encoder in place of its original embedding layer. This modification allows for more effective semantic learning from assembly code, improving the model’s ability to capture the intricacies of assembly-to-source code mapping. The architecture of DECodeT5 not only accelerates convergence during end-to-end decompilation tasks but also supports rapid adaptation across different compilers and instruction set architectures (ISAs) by swapping out the encoder module as needed. Our experimental evaluation on the HumanEval and Exebench benchmarks reveals that DECodeT5 significantly improves semantic recovery accuracy, outperforming Ghidra by over 83% and LLM4Decompile by more than 43% in terms of execution recovery. Furthermore, DECodeT5 maintains a compact model size of only 360M parameters, providing inference speeds that are twice as fast as LLM4Decompile. These results underscore DECodeT5’s suitability for deployment in resource-constrained environments and its flexibility in adapting to real-world reverse engineering scenarios, offering a practical solution for modern firmware analysis and security tasks.

The precision of binary disassembly is crucial for understanding program behavior in reverse engineering. However, existing disassembly tools struggle to accurately identify function boundaries when binary files are stripped, especially on reduced instruction set computer (RISC) architectures like ARM and MIPS. Additionally, disassembly tools often lack support for newer versions or less used architectures. In this paper, we propose UDA, a unified pre-training disassembly framework with multi-architecture support. Unified pre-training on both machine code and assembly code allows assembly code to help the semantic learning of machine code, thereby addressing the inherent semantic deficiencies of machine code. By extracting enhanced semantic features from machine code, UDA can not only more accurately identify function boundaries but also recover assembly code without relying on existing disassembly tools. UDA has been rigorously evaluated against both regular and obfuscated binary files. The results show that UDA improves the F1 score for function boundary detection by 0.9%, 7.9%, and 8.1% on x86, ARM, and MIPS architectures, respectively, compared to the state-of-the-art disassemblers. Additionally, it demonstrates strong robustness against unseen obfuscated binaries.

Model Inversion Attacks (MIAs) against face recognition systems aim to reconstruct facial images of specific individuals from the recognition models. Existing MIA approaches commonly optimize the latent variables of Generative Adversarial Networks (GANs) iteratively, which can result in non-smooth optimizations due to the complexity and entanglement of latent space. Furthermore, the optimization guided by the target model’s gradients may generate high-confidence images with poor perceptual similarity to the target class. This paper introduces a novel perspective by reformulating the inversion attack as a conditional data distribution learning task. Based on this, we propose a Feature-Guided Model Inversion Attack (FGMIA), which learns the facial data distribution and integrates feature guidance as a conditional signal. Specifically, we treat the deconstructed target model as a feature encoder, which provides guidance during the training of a specialized feature-guided diffusion model. During the attack, feature encodings implicit in the target model are extracted and utilized to guide the reconstruction of private data. Extensive experiments demonstrate that FGMIA accurately reconstructs private data from face recognition models and significantly improves evaluation accuracy and perceptual similarity compared to state-of-the-art methods while maintaining comparable target confidence scores. Our code is available at https://github.com/MMCTTT/FGMIA_codes.

Federated learning (FL) is a distributed machine learning paradigm ensuring data privacy. However, the statistical heterogeneity poses a challenge to building a single model that can perform well across all clients’ data distributions. Personalized FL (PFL) has emerged as a solution to mitigate the impact of statistical heterogeneity by training separate models for various target data distributions. A crucial aspect of PFL is to utilize additional information in the federated system to assist in training personalized models. In this article, we introduce a novel PFL approach that leverages a multiview geometry structure (GPFL). GPFL formulates an optimization problem to determine the correlation weights among clients by utilizing the geometry structure composed of client gradient updates. It further builds information carriers that facilitate personalized training based on the weights. To accurately capture the correlations, we use both L2 distance and cosine similarity views to depict geometric similarity. In federated training, the discrepancy in timeliness and tendency to overfit to local data in gradient updates cause the geometric similarity of such updates to inadequately reflect the client relationships. Therefore, GPFL employs representative gradients extracted from the client’s historical gradients to infer the correlation weights. Experimental results on four data sets, convex and nonconvex objectives, and two FL settings demonstrate that our method outperforms several PFL methods.

Recently, it has been shown that adversaries can reconstruct images from SIFT features through reverse attacks. However, the images reconstructed by existing reverse attack methods suffer from information loss and are unable to sufficiently reveal the private contents of the original images. In this paper, a two-stage deep reverse attack model called Coarse-to-Fine Generative Adversarial Network (CFGAN) is proposed to more deeply explore the information in SIFT features and further demonstrate the risk of privacy leakage associated with SIFT features. Specifically, the proposed model consists of two sub-networks, namely coarse net and fine net. The coarse net is developed to restore coarse images using SIFT features, while the fine net is responsible for refining the coarse images to obtain better reconstruction results. To effectively leverage the information contained in SIFT features, an efficient fusion strategy based on the AdaIN operation is designed in the fine net. Additionally, we introduce a new loss function called sift loss that enhances the color fidelity of reconstructed images. Extensive experiments conducted on various datasets verify that the proposed CFGAN performs favorably against state-of-the-art methods. The reconstructed images exhibit better visual quality, less texture distortion, and higher color fidelity. Source code is available at https://github.com/HITLiXincodes/CFGAN.

The widespread collection and analysis of private speech signals have become increasingly prevalent, raising significant privacy concerns. To protect speech signals from unauthorized analysis, adversarial attack methods for deceiving speaker recognition models have been proposed. While a few of these methods are specifically designed for real-time protection of speech signals, they introduce significant delays that can severely impact speech communication when applied to streaming speech data. In this paper, we present a novel approach that aims to offer real-time protection for speech signals without delays. By utilizing observed data only, we generate initial adversarial seed perturbations and refine them to obtain the necessary adversarial perturbations predicted for adjacent unobserved signals. This refinement process is conducted via a proposed model called PAPG. On the basis of perturbation prediction, we develop a streaming audio processing framework that generates perturbations in synchronization with the playback of the original signal, effectively eliminating delays. The experimental results demonstrate that under the proposed attack, the average Top-1 accuracy of various advanced speaker recognition methods is reduced by 89%, and the average equal error rate (EER) increases to 36%. Remarkably, these results are achieved without delays while maintaining superior perceptual quality.

Federated learning (FL) is a distributed machine learning framework that allows resource-constrained clients to train a global model jointly without compromising data privacy. Although FL is widely adopted, high degrees of systems and statistical heterogeneity are still two main challenges, which leads to potential divergence and nonconvergence. Clustered FL handles the problem of statistical heterogeneity straightly by discovering the geometric structure of clients with various data generation distributions and getting multiple global models. The number of clusters contains prior knowledge about the clustering structure and has a significant impact on the performance of clustered FL methods. Existing clustered FL methods are inadequate for adaptively inferring the optimal number of clusters in environments with high systems' heterogeneity. To address this issue, we propose an iterative clustered FL (ICFL) framework in which the server dynamically discovers the clustering structure by successively performing incremental clustering and clustering in one iteration. We focus on the average connectivity within each cluster and give incremental clustering and clustering methods that are compatible with ICFL based on mathematical analysis. We evaluate ICFL in experiments on high degrees of systems and statistical heterogeneity, multiple datasets, and convex and nonconvex objectives. Experimental results verify our theoretical analysis and show that ICFL outperforms several clustered FL baseline methods.

With the widespread use of automated speech recognition (ASR) systems in modern consumer devices, attack against ASR systems have become an attractive topic in recent years. Although related white-box attack methods have achieved remarkable success in fooling neural networks, they rely heavily on obtaining full access to the details of the target models. Due to the lack of prior knowledge of the victim model and the inefficiency in utilizing query results, most of the existing black-box attack methods for ASR systems are query-intensive. In this paper, we propose a new black-box attack called the Monte Carlo gradient sign attack (MGSA) to generate adversarial audio samples with substantially fewer queries. It updates an original sample based on the elements obtained by a Monte Carlo tree search. We attribute its high query efficiency to the effective utilization of the dominant gradient phenomenon, which refers to the fact that only a few elements of each origin sample have significant effect on the output of ASR systems. Extensive experiments are performed to evaluate the efficiency of MGSA and the stealthiness of the generated adversarial examples on the DeepSpeech system. The experimental results show that MGSA achieves 98% and 99% attack success rates on the LibriSpeech and Mozilla Common Voice datasets, respectively. Compared with the state-of-the-art methods, the average number of queries is reduced by 27% and the signal-to-noise ratio is increased by 31%.

In the absence of prior knowledge, it is a challenge to determine the number of protocol types in frames, which are completely unknown. These frames might be mixed with multiple protocols from the data link layer to the application layer. In this study, the authors combine the spectral clustering algorithm with a method of determining the number of protocol frame types, and further more design the refinement clustering of different hierarchical protocols based on the eigenvectors of the Laplace matrix. They use three clustering validity indices, which are Calinski–Harabasz index, Davies–Bouldinn index and Silhouette index, to quantify the clustering effect in order to calculate the number of protocol types. Extensive experiments on several open datasets obtain relatively satisfying results without prior knowledge and demonstrate the significant advantages of their methods clearly.