Machine learning has the potential to revolutionize medical image classification. However, machine learning requires large medical datasets to improve accuracy, which will compromise patient privacy. Federated learning is a promising technique that protects patient privacy and improves the accuracy of medical image classification. Unfortunately, current research shows that federated learning faces the risk of privacy leakage. In this paper, we propose a privacy-preserving federal learning scheme via hybrid differential privacy for medical image classification (FHDM). Specifically, we construct a local hybrid differential privacy algorithm (LHDP) against patients' privacy leakage. This hybrid algorithm combines Gaussian and Laplace differential privacy without enlarging the privacy budget. We prove that the algorithm applies to the model parameter. Moreover, we design loss optimization and global optimization strategies on the algorithm to achieve higher accuracy in medical image classification. Finally, we validate FHDM in terms of privacy-preserving and model accuracy on real datasets. Experiments show that FHDM effectively protects privacy and improves the average accuracy by 9.60% compared to previous differential privacy schemes with the same medical image dataset and privacy budget.

The anomaly identification of vibration monitoring signals for rotating machinery supports fault diagnosis. The abnormal representation of the monitoring signals for rotating machinery faults is a long signal with periodic characteristics and multipoint correlations. Deep learning methods can effectively distinguish various types of faults by learning the interpoint representations or the correlation patterns among sampling points in the monitoring data. However, under the influence of noise, the fault features are disrupted, resulting in a deterioration of the fault identification effect. We noticed that the amplitude of the signal is superimposed with noise, which affects the probability distribution of the amplitude of the original signal. To solve this problem, we proposed a novel deep learning model named variational spectrum-attention transformer (VSAT) with a spectrum attention mechanism. It uses frequency-domain projection to transform the signal amplitude distribution into point distributions at multiple frequencies. Then, an exponential function is employed to increase the distances between feature points. An attention branch for learning the statistical distribution patterns of amplitudes at different frequencies is constructed. It uses variational method to learn the distribution parameters of the frequency-domain amplitudes of monitored sequential data. An experimental analysis of weight grad-CAM maps demonstrates that, compared to extracting features from frequency-domain data by transformer model, the new method captures different distribution features. The verification and comparison results on 3 bearings and 1 gearbox dataset show that the new model not only accelerates the model convergence, but also improves the accuracy of anomaly classification, and has excellent performance in multiclassification and antinoise performance.

Modern security methods use address space layout randomization (ASLR) to defend against code reuse attacks (CRAs). However, code probing can still obtain the content and address of the code through code probing. Code probing invalidates the widely used ASLR methods, causing researchers to lose confidence in them. On the contrary, we believe the ASLR is still effective, if it has anti-probing capability. To enhance the anti-probing capability of ASLR and defense CRAs, this paper proposes an anti-probing method MProbe. First, it detects the code probing activities of attackers, including address probing and content probing. Next, the execution permission of the probed code will be de-enabled in the original address space. At the same time, the equivalent code block in a random address space will replace the probed code. Finally, new security strategies are used to prevent the probed code blocks from being used as gadgets. Experiments and analysis show that MProbe has a good defense effect against CRAs based on code probing, and only introduces less than 3% performance overhead to the operating system (OS).

“Out-of-box” methods can provide strict isolation between the security tool and the target virtual machine (TVM), so they have strong anti-jamming abilities. Current monitoring methods are mostly based on memory snapshots at a time point or a fixed period. Due to the randomness of processes and the discontinuity of monitoring methods, methods based on time point or fixed period snapshots that security tools use to monitor processes may result in a high leakage rate. The real-time monitoring method can be used to resolve this problem. However, all current real-time monitoring methods require a ready monitoring set, and their monitoring range is strictly limited. They are only effective for the inherent objects within the ready set and ineffective for the random processes. This paper presents RMVP, a real-time monitoring method to monitor random process in TVM. First, RMVP monitors process switch by capturing the switch of kernel stack base addresses outside TVM in real time. Then it extracts raw memory of the current process through memory mapping technology that adopts caching mechanism and multi-task concurrent execution strategy. Finally, RMVP translates raw memory into high-level semantics according to a semantic knowledge base constructed offline. RMVP can monitor random processes in real time and overcome the challenge of process randomness. The experimental results show that the capture rate of random process is over 95% and the capture delay is in the range of 2.3ms ~ 3.3ms. In addition, RMVP is especially effective for detecting the processes hidden by rootkits.