ABSTRACT In the context of federated learning, there are certain doubts about the credibility of cloud servers as third parties, as they have the potential to extract sensitive information from participants' local data from gradients. In addition, cloud servers may even resort to forging aggregation results, leading to the destruction of the global model and successfully avoiding detection mechanisms. Therefore, in building a secure federated learning system, it is crucial to ensure the privacy and aggregation correctness of upload gradients. This article proposes a secure and efficient federated learning privacy protection scheme TreePPFL (Tree Privacy Protection Federated Learning) based on hop‐by‐hop communication verifiability. By using single concealment technology to encrypt model parameters, the privacy of the uploaded gradient is protected. At the same time, a bidirectional verification scheme was designed, which applies a homomorphic hash algorithm to enable the cloud server to verify the legitimacy of the client, while also enabling the client to verify the aggregation correctness of the cloud server. The client transmits information through hop‐by‐hop communication, improving the training and verification efficiency of the cloud server and the entire federated learning system. The security of the scheme was verified through security analysis. The empirical experiment used two publicly available datasets, MNIST and CIFAR‐100, and compared them in iid and noniid scenarios. The results showed that the TreePPFL scheme exhibited superior performance compared to other schemes.

In the robustness verification of neural networks, formal methods have been used to give deterministic guarantees for neural networks. However, recent studies have found that the verification method of single-neuron relaxation in this field has an inherent convex barrier that affects its verification capability. To address this problem, we propose a new verification method by combining dual-neuron relaxation and linear programming. This method captures the dependencies between different neurons in the same hidden layer by adding a two-neuron joint constraint to the linear programming model, thus overcoming the convex barrier problem caused by relaxation for only a single neuron. Our method avoids the combination of exponential inequality constraints and can be computed in polynomial time. Experimental results show that we can obtain tighter bounds and achieve more accurate verification than single-neuron relaxation methods.

The verification techniques based on formal methods can provide deterministic guarantees for the robustness of Deep Neural Networks(DNNS). However, the enormous scale of DNNS makes the application of such methods in this field a huge challenge. To address this problem, this study proposes a single-sample sub-network pruning method, which can identify redundant nodes by combining neuron coverage and the symbolic interval propagation method to reduce the network verification scale. In addition, to solve the problem of too many sub-networks to be pruned, according to the similarity of neuron coverage between samples, we propose a corresponding clustering algorithm to establish sub-networks for different categories of samples to improve the verification efficiency. We combine the MIPverify verification tool to validate the above method. Experiments show that the sub-networks can give the same robust validation results and similar robustness bounds as the original network, while greatly reducing the validation time and network size.

Abstract Self-driving cars over the last decade have achieved significant progress like driving millions of miles without any human intervention. However, behavioral safety in applying deep-neural-network-based (DNN based) systems for self-driving cars could not be guaranteed. Several real-world accidents involving self-driving cars have already happened, some of which have led to fatal collisions. In this paper, we present a novel and automated technique for verifying steering angle safety for self-driving cars. The technique is based on deep learning verification (DLV), which is an automated verification framework for safety of image classification neural networks. We extend DLV by leveraging neuron coverage and slack relationship to solve the judgement problem of predicted behaviors, and thus, to achieve verification of steering angle safety for self-driving cars. We evaluate our technique on the NVIDIA’s end-to-end self-driving architecture, which is a crucial ingredient in many modern self-driving cars. Experimental results show that our technique can successfully find adversarial misclassifications (i.e., incorrect steering decisions) within given regions if they exist. Therefore, we can achieve safety verification (if no misclassification is found for all DNN layers, in which case the network can be said to be stable or reliable w.r.t. steering decisions) or falsification (in which case the adversarial examples can be used to fine-tune the network).

With the increasing popularity of smartphones, the mobile security issues have become serious, and more and more malware has been found. Android applications are often used to handle sensitive information, thus they have become the main targets of malware attacks. In order to efficiently detect Android malware, in this paper, we present a multi-strategy combination framework. We use five types of static features to characterize Android applications from multiple aspects. To improve the classification accuracy and reduce the overfitting of the framework, we use three filter-based feature selection methods to identify the most informative top-k features. Then we input the applications represented by the feature subsets into five classification algorithms to build classifiers. Finally, we predict the classification results by hard voting or soft voting. We have performed many experiments in a well-marked dataset consisting of 41,155 samples. The experimental results show that our approach can achieve over 98% in accuracy, precision, recall and F-score. Compared with other existing methods, our approach has the best malware detection rate of 98.75%.

The real-time interaction between embedded software and its external environment is conducted through the interrupt mechanism. Since the interrupt request is random and responds according to priority, the execution of embedded software is non-sequential, which leads to the non-deterministic software behaviors. If these non-deterministic behaviors can be quantitatively pre-analyzed during the software design phase, the reliability of embedded software can be improved effectively. In this paper, we first provide an embedded software behavior model based on extended deterministic and stochastic Petri nets (EDSPN). Through EDSPN, the interrupt behavior of embedded software can be effectively modeled. Then we put forward a probabilistic model checking method of Continuous Stochastic Logic (CSL) for EDSPN to analyze embedded software behavior. For alleviating the state explosion problem, the above method uses the bounded model checking (BMC) technique. We present the model checking methods and the probability metric calculation methods for CSL operators under bounded semantics. Finally, by analyzing the EDSPN model of embedded software with multiple interrupts, we compare the analytical capabilities of BMC method and non-BMC method. The experiment shows that when the state space of EDSPN is large and is hard to calculate, the bounded checking algorithm can be used to approximate the software behavior. The conclusions obtained are helpful to understand the properties to be verified.

Craig interpolation has emerged as an effective approximation method and can be widely applied in hardware and software model checking. Since the quality of interpolants can critically affect the success and failure, or convergence and divergence of model checking, researchers have put forward a novel and flexible interpolation abstraction-based technique to guide the computation of promising interpolants. In this technique, abstraction lattice is constructed to arrange families of interpolation abstraction for improving the quality of resulting interpolants. However, the original search strategy to explore an abstraction lattice is not efficient when abstraction lattice enlarges and the elapsed time to perform multiple search on the same abstraction lattice is obviously distinct for many problems. In this paper, in order to alleviate these problems, we propose a top-down search space pruning-based algorithm to search the abstraction lattice and implement this algorithm in the well-known model checker Eldarica. We conduct experiments on 179 benchmarks to compare our algorithm respectively against the original search algorithm in Eldarica and the state-of-the-art SMT solver Z3. The experimental results show that our algorithm performs much better in the sense that it is more efficient than Eldarica for most of the benchmarks and it can solve much more benchmarks than Z3.