Gongcheng Kexue Yu Jishu/Advanced Engineering Science

In digital signal processing, Finite Impulse Response (FIR) filters are fundamental components that govern behavior by specific coefficients. Protecting these coefficients from adversarial discovery is crucial, and prior techniques have introduced obfuscation through decoys or key bits. However, these methods remain vulnerable to query attacks capable of revealing secret keys. This work introduces a novel hybrid protection technique that combines hardware obfuscation and logic locking, utilizing a point function to secure parallel direct and transposed forms of FIR filters. Additionally, a Proportional-Integral-Derivative (PID) controller is integrated to enhance system performance and resilience. An innovative unlock function is also developed to ensure controlled access to the protected coefficients. Experimental results demonstrate that the proposed technique offers superior security, protecting against prominent query-based and logic attacks while maintaining competitive hardware complexity. The protected multiplier blocks and FIR filters, including the integrated PID controller, show enhanced security and efficiency compared to traditional methods.

The explosive growth of data generated on social media platforms like Twitter presents both opportunities and challenges for real-time anomaly detection. Traditional approaches struggle to scale with the velocity, volume, and variety of such data. This paper proposes a scalable framework for anomaly detection using stream mining techniques built on Apache Spark and its machine learning library, MLlib. The system is designed to process high-throughput tweet streams in real time, detect anomalous patterns, and evaluate the performance of various anomaly detection algorithms including Streaming K-Means and Isolation Forests. Twitter data is used as a dynamic, high-velocity input source to simulate real-world streaming environments. The framework leverages Spark Streaming for real-time ingestion and distributed processing, while MLlib enables scalable model training and inference. Comprehensive performance analysis is conducted using key metrics: speedup to measure parallel processing efficiency across Spark clusters, throughput to quantify the volume of tweets processed per second, and accuracy to evaluate detection effectiveness. Experimental results demonstrate that the proposed framework can achieve near-linear speedup, high throughput, and competitive accuracy in identifying tweet-based anomalies, such as sudden spikes in sentiment, bot-like behavior, or misinformation bursts. The system proves to be robust, scalable, and suitable for deployment in domains requiring low-latency insights from continuous data streams.

The objective of this study is to develop a robust and scalable framework for detecting fake product reviews in large-scale e-commerce platforms using Graph Neural Networks (GNNs). The research aims to improve the accuracy of review authenticity classification by modeling user-product-review interactions as graph structures, capturing relational and behavioral patterns beyond traditional feature-based methods.The proposed approach constructs a heterogeneous graph from an e-commerce dataset where nodes represent users, products, and reviews, and edges encode their relationships. We apply a Graph Neural Network model that leverages both structural and semantic features extracted from the graph. Attention mechanisms are incorporated to focus on influential connections. The model is trained and evaluated on publicly available datasets such as Amazon and Yelp, with preprocessing steps including review vectorization, graph formation, and normalization. Benchmark comparisons are made against conventional machine learning classifiers and deep learning models. Experimental results demonstrate that the GNN-based model significantly outperforms baseline methods, achieving a higher F1-score, precision, and recall across multiple datasets. The graph-based approach exhibits superior ability to capture deceptive behavior patterns and uncover collusive review activities. The model also shows scalability when applied to datasets with millions of reviews.