Continuously publishing histograms in data streams is crucial to many real-time applications, as it provides not only critical statistical information, but also reduces privacy leaking risk. As the importance of elements usually decreases over time in data streams, in this paper we model a data stream by a sequence of weighted sliding windows, and then study how to publish histograms over these windows continuously. The existing literature can hardly solve this problem in a real-time way, because they need to buffer all elements in each sliding window, resulting in high computational overhead and prohibitive storage burden. In this paper, we overcome this drawback by proposing an online algorithm denoted by Efficient Streaming Histogram Publishing (ESHP) to continuously publish histograms over weighted sliding windows. Specifically, our method first creates a novel sketching structure, called Approximate-Estimate Sketch (AESketch), to maintain the counting information of each histogram interval at every time instance; then, it creates histograms that satisfy the differential privacy requirement by smartly adding appropriate noise values into the sketching structure. Extensive experimental results and rigorous theoretical analysis demonstrate that the ESHP method can offer equivalent data utility with significantly lower computational overhead and storage costs when compared to other existing methods.

Deep neural networks are commonly used in computer vision tasks, but they are vulnerable to adversarial samples, resulting in poor recognition accuracy. Although traditional algorithms that craft adversarial samples have been effective in attacking classification models, the attacking performance degrades when facing object detection models with more complex structures. To address this issue better, in this paper we first analyze the mechanism of multi-scale feature extraction of object detection models, and then by constructing the object feature-wise attention module and the perturbation extraction module, a novel adversarial sample generation algorithm for attacking detection models is proposed. Specifically, in the first module, based on the multi-scale feature map, we reduce the range of perturbation and improve the stealthiness of adversarial samples by computing the noise distribution in the object region. Then in the second module, we feed the noise distribution into the generative adversarial networks to generate adversarial perturbation with strong attack transferability. By doing so, the proposed approach possesses the ability to better confuse the judgment of detection models. Experiments carried out on the DroneVehicle dataset show that our method is computationally efficient and works well in attacking detection models measured by qualitative analysis and quantitative analysis.

The problem of imbalanced data classification learning has received much attention. Conventional classification algorithms are susceptible to data skew to favor majority samples and ignore minority samples. Majority weighted minority oversampling technique (MWMOTE) is an effective approach to solve this problem, however, it may suffer from the shortcomings of inadequate noise filtering and synthesizing the same samples as the original minority data. To this end, we propose an improved MWMOTE method named joint sample position based noise filtering and mean shift clustering (SPMSC) to solve these problems. Firstly, in order to effectively eliminate the effect of noisy samples, SPMSC uses a new noise filtering mechanism to determine whether a minority sample is noisy or not based on its position and distribution relative to the majority sample. Note that MWMOTE may generate duplicate samples, we then employ the mean shift algorithm to cluster minority samples to reduce synthetic replicate samples. Finally, data cleaning is performed on the processed data to further eliminate class overlap. Experiments on extensive benchmark datasets demonstrate the effectiveness of SPMSC compared with other sampling methods.