In the era of rapid digital communication and exponential data growth, ensuring the confidentiality and integrity of transmitted information is paramount. This paper addresses the need for robust and efficient cryptographic solutions by optimizing authenticated encryption schemes. We present a comprehensive study on enhancing the performance of ChaCha20 and Blake3, focusing on high-performance, parallel implementations suited for large-scale data handling. Our ChaCha20 implementation leverages vectorization and parallelization to achieve up to 55x speedup, surpassing existing libraries like libsodium. For Blake3, we explore both divide-and-conquer and stack-based approaches, demonstrating a scalable improvement in performance. The divide-and-conquer method achieves a peak throughput of 39-50 GB/s, while the stack-based approach shows substantial efficiency gains. This work contributes valuable insights into the optimization of authenticated encryption, highlighting significant advancements in cryptographic performance for contemporary applications.

This paper introduces DetectBERT, a cuttingedge deep learning framework designed for statement-level vulnerability detection (SVD) in source code, with a focus on Python. In contrast to conventional vulnerability detection methods that rely on graph-based structures like Control Flow Graphs (CFG) or Data Dependency Graphs (DDG), DetectBERT utilizes the BERT transformer encoder to learn contextual relationships between code statements directly from the text. The model consists of two BERT-based layers: one dedicated to feature extraction and the other for classification. By eliminating the need for graph-based representations, DetectBERT offers significant advantages, such as handling the dynamic nature of programming languages and managing variations in code across different versions. This approach also reduces reliance on language-specific graph generation tools, enhancing its flexibility and scalability. Experimental results on publicly available datasets show that DetectBERT outperforms traditional graph-based methods in vulnerability detection. Moreover, the paper introduces a novel preprocessing pipeline that labels Python code by analyzing commit history, further boosting the model's accuracy and resilience. DetectBERT's design is extendable to other programming languages, and the complete implementation is made available to foster continued research and practical applications in the domain of software security.

This paper presents the integration of Intel Software Guard Extensions (SGX) with the Confidential Measurement Control (CMC) framework to improve the process of remote attestation in Trusted Execution Environments (TEEs). While existing solutions like Intel's SGX provide hardware-based security, the CMC framework enhances the flexibility and reliability of report generation and validation. Our approach focuses on incorporating the CMC as a program library within SGX enclaves, addressing challenges related to software manifests, CRL updates, and measurement verification. The report generation and verification processes are detailed, alongside an analysis of performance and security. We explore various verification steps, from certificate chain validation to integrity checks on hardwarebased measurements, and highlight optimizations that reduce overhead during attestation. Finally, the paper discusses the potential integration of emerging cryptographic techniques like homomorphic encryption and multi-party computation for future scalability and confidentiality improvements in remote attestation.

Modern modular service ecosystems rely on microservices for scalability and flexibility, yet performance constraints such as bottlenecks and resource contention remain critical challenges. This research presents a novel insight-driven framework for detecting and mitigating performance anomalies in microservices-based applications. By leveraging graph-based learning models and multi-source telemetry analysis, our approach effectively identifies independent, dependent, and cascading bottlenecks, facilitating timely intervention. We introduce and evaluate GAMMA, an advanced anomaly detection and localization model that integrates attention-based graph convolution networks with a multi-expert learning mechanism. Empirical evaluations on large-scale datasets demonstrate that GAMMA significantly outperforms existing methods in detecting and diagnosing multiple bottlenecks, achieving an F1 score of 0.92 for anomaly detection and 0.89 for bottleneck localization. The proposed framework enhances performance management in modular service architectures, ensuring resilient and adaptive cloud-native applications.