Search for: All records

Given the high cost of large language model (LLM) training from scratch, safeguarding LLM intellectual property (IP) becomes increasingly crucial. As the standard paradigm for IP ownership verification, LLM fingerprinting thus plays a vital role in addressing this challenge. Existing LLM fingerprinting methods verify ownership by extracting or injecting model-specific features. However, they overlook potential attacks during the verification process, leaving them ineffective when the model thief fully controls the LLM's inference process. In such settings, attackers may share prompt-response pairs to enable fingerprint unlearning, or manipulate outputs to evade exact-match verification. We propose iSeal, the first fingerprinting method designed for reliable verification when the model thief controls the suspected LLM in an end-to-end manner. It injects unique features into both the model and an external module, reinforced by an error-correction mechanism and a similarity-based verification strategy. These components are resistant to verification-time attacks, including collusion-based fingerprint unlearning and response manipulation, backed by both theoretical analysis and empirical results. iSeal achieves 100% Fingerprint Success Rate (FSR) on 12 LLMs against more than 10 attacks, while baselines fail under unlearning and response manipulations. 

Breadth-First Search (BFS) is a fundamental graph traversal algorithm in a level-by-level pattern. It has been widely used in real-world applications, such as social network analysis, scientific computing, and web crawling. However, achieving high performance for BFS on large-scale graphs remains a challenging task due to irregular memory access patterns, diverse graph structures, and the necessity for efficient parallelization. This paper addresses these challenges by designing a highly optimized parallel BFS implementation based on the top-down and bottom-up traversal strategies. It further integrates several key innovations, including graph typea-ware computation strategy selection, graph pruning, twolevel bottom-up, and efficient parallel implementation. We evaluate our method on 11 diverse graphs in terms of size, diameter, and density. On a CPU server with 48 threads, our method achieves an average speedup of 9.5x over the serial BFS implementation. Also, on a synthetic dense graph, our method processes 9.3 billion edges per second, showing its efficiency in dense graph traversal. 

The single-source shortest path (SSSP) problem is essential in graph theory with applications in navigation, biology, social networks, and traffic analysis. The  -Stepping algorithm enhances parallelism by grouping vertices into "buckets" based on their tentative distances. However, its performance depends on values and graph properties. This paper introduces an adaptive parallel Delta-Stepping implementation with three innovations: neighbor reordering, bucket fusion, and graph type-aware selection. Tested on 11 diverse graphs, it achieves an average 7.1× speedup over serial Dijkstra’s algorithm on a 48-thread CPU server.