Deep learning-based log anomaly detection demands extensive labeled data, posing significant challenges for emerging systems with limited logs. Transfer learning mitigates this issue by leveraging knowledge from data-rich source domains, enabling effective adaptation to data-scarce target domains. The challenge in recent cross-domain research lies in the joint optimization of knowledge transfer efficacy and model generalization capability. To address these limitations, we propose LogCADA, a novel logarithmic anomaly detection framework based on transfer learning, which can obtain effective common features through double-layer adversarial training, and distinguish common features and unique features between different domains through multi-source domain contrast alignment to achieve better knowledge transfer.The results demonstrate that our method can be adapted from dual source to single target domain and effectively overcome the inherent limitations of traditional cross-domain anomaly detection methods, yielding significant practical value for real-world log analysis scenarios.

Log anomaly detection is critical for maintaining system reliability and observability in complex cloud and microservice environments. However, existing methods often remain limited to binary classification, struggle to adapt to dynamic log patterns, and suffer from semantic loss due to log parsing. To address these challenges, we propose GLog, an end-to-end framework that enables dynamic anomaly type prediction without requiring manual type labels. GLog first fine-tunes instruction-tuned large language models using normal/abnormal labels to achieve high-accuracy anomaly detection on raw, unparsed log sequences. It then clusters the detected anomalies to automatically generate pseudo anomaly type labels and descriptions, which are further used for second-stage fine-tuning, enabling the model to predict specific anomaly types with interpretable outputs. By leveraging full log semantics and dynamically updating its anomaly type repository, GLog reduces manual annotation costs and adapts to evolving system behaviors in large-scale environments.

Anomaly detection in HTTP traffic is critical for securing web applications against evolving cyber threats. We propose a semi-supervised method that combines domain-specific language modeling with sequence reconstruction to identify anomalies in HTTP requests. Our approach leverages only benign traffic for training and uses reconstruction errors for detecting malicious activity. It achieves a strong balance between precision and recall while maintaining low computational requirements, making it suitable for real-time and edge deployments. Extensive evaluations on three public HTTP datasets show that our method outperforms traditional baselines and fine-tuned BERT models, with an F1-score of 0.92 and AUC of 0.96. We also introduce a simple interpretability mechanism by attributing anomalies to token-level reconstruction errors, providing insights into detected threats. The proposed solution is scalable, lightweight, and effective across diverse attack scenarios without requiring large labeled datasets.

System logs record the system’s status and application behavior, providing support for various system management and diagnostic tasks. However, existing methods for log anomaly detection face several challenges, including limitations in recognizing current types of anomalous logs and difficulties in performing online incremental updates to the anomaly detection models. To address these challenges, this paper introduces DRLLog, which applies Deep Reinforcement Learning (DRL) networks to detect anomalous events. DRLLog uses Deep Q Network (DQN) as the agent, with log entries serving as reward signals. By interacting with the environment generated from log data and adopting various action behaviors, it aims to maximize the reward value obtained as feedback. Through this approach, DRLLog achieves learning from historical log data and perception of the current environment, enabling continuous learning and adaptation to different log sequence patterns. Additionally, DRLLog introduces low-rank adaptation by using two low-rank parameter matrices in the fully connected layer of the DQN to represent changes in its weight matrix. During online model learning, only low-rank parameter matrices of the model are updated, effectively reducing the model’s overhead. Furthermore, DRLLog introduces focal loss to focus more on learning the features of anomalous logs, effectively addressing the issue of imbalanced quantities between normal and anomalous logs. We evaluated the performance on widely used log datasets, including HDFS, BGL and ThunderBird, showing an average improvement of 3% in F1-Score compared to baseline methods. During online model learning, DRLLog achieves an average reduction of 90% in parameter count and a significant decrease in training and testing time as well.

Image compression has become an important task for reducing storage and transmission costs. However, recent models for learned image compression have been developed to increase the network’s number of layers and channels to achieve better visual effects. This resulted in higher computing and memory resources, making deploying the model on compute-constrained platforms such as wireless devices impractical. In this paper, we propose a lightweight autoencoder with hierarchical priors. The lightweight autoencoder reduces the model’s parameter size and calculation amount based on ensuring high fidelity and low bit rates of the image. Simulation results indicate the proposed model yields a smaller size: the parameters are reduced by 81.66%, and the calculation amount is reduced by 94.7% over the benchmark. Besides, the proposed model results in a speed improvement of 200 times. At the same time, our model achieves nearly the same performance as the baseline on MS-SSIM and LPIPS distortion metrics.

As an alternative implementation of Slepian-Wolf coding, distributed arithmetic coding is very competitive in short and medium block lengths. The existing distributed arithmetic coding decoder uses the maximum a posteriori metric and the M-algorithm to select the decoding sequence by considering the prior information. Since the prior probability distribution has been explored by the encoding process via the model stage and retained in the codeword, we propose a modified decoding metric ignoring the prior information. The reliability of the decoding paths is only determined by the correlation between the side information and the input source. Simulation results show that the modified metric can significantly reduce the decoding error.

Facial landmark localization on images with occlusions is an important and challenging task in many visual applications. Recently, the cascaded pose regression has attracted increasing attention, since it achieved superior performance in terms of facial landmark localization under occlusions. However, such approach is sensitive to initialization, where an improper initialization will decrease the performance sharply. In this paper, we propose a novel initialization method to get a robust initial shape by analysing correlation of Local Binary Patterns (LBP) histograms between the estimated face and training faces. The shape of the training face that is most correlated with the estimated face, will be selected as the initialization for the regression. The selected shape is closer to the real shape of the estimated face, which makes the landmark localization more accurate. Besides, in order to make the initial shape more robust to occlusions, we propose a boosted smart restarts technique by checking location and occlusion jointly instead of checking location only. We show that the proposed method significantly improves performance over existing landmark localization methods on the challenging dataset of COFW. The experimental results demonstrate that the proposed method reduces error by 11.9% and failure cases by 20.8% on COFW dataset. Moreover, it detects face occlusions with 85/40% precision/recall.