AN ATTENTION-ENHANCED DEEP LEARNING FRAMEWORK FOR PREDICTIVE SOFTWARE LOOP VULNERABILITY DETECTION, FAULT LOCALIZATION AND OPTIMIZATION THROUGH SOURCE-CODE BEHAVIORAL ANALYSIS

Abstract

Software loops are critical components of many applications, yet their repeated operations, complex control flow, and changing program states can conceal vulnerabilities and make faults difficult to locate. This paper presents an attention-enhanced deep learning framework for predicting software loop vulnerabilities, localizing faulty statements, and identifying safe optimization opportunities through source-code behavioral analysis. The framework integrates abstract syntax trees, control-flow and data-flow graphs, loop execution traces, variable-update patterns, and static complexity measures. A graph encoder learns structural relationships, a bidirectional sequence model captures execution behavior, and a cross-attention mechanism connects suspicious operations with their program context. Three task-specific outputs provide vulnerability predictions, ranked fault locations, and optimization recommendations. Recommended changes undergo compilation, regression testing, and semantic-equivalence checks before acceptance. The study uses 120,000 loops extracted from 9,600 open-source C and C++ projects. The dataset includes 18,000 vulnerability-labelled loops, 24,000 fault-localization instances linked to developer-confirmed fixes, and 30,000 loops with validated optimization opportunities. Vulnerability categories include out-of-bounds access, integer overflow, nontermination, unsafe pointer updates, and resource misuse. Following duplicate removal, projects are assigned to training, validation, and independent test sets in a 70:15:15 ratio to limit information leakage. The framework is compared with rule-based analysis, CodeBERT, a graph neural network, and a transformer without behavioral attention. On the held-out test set, the proposed framework achieves 94.2% vulnerability-detection accuracy, 92.1% macro-F1, 91.4% precision, and 92.8% recall. It places the faulty statement among the top three ranked locations in 88.6% of cases, compared with 76.3% for the strongest baseline. Across eligible benchmarks, validated recommendations reduce median loop execution time by 17.5% and peak memory use by 9.2%, with no observed regression-test failures. Ablation experiments show that removing behavioral features reduces macro-F1 by 4.6 percentage points, while removing attention reduces top-three localization accuracy by 5.1 points. A manual review of 200 flagged loops further indicates that attention maps can help reviewers inspect relevant data dependencies. These findings support the combined use of structural and behavioral information for more accurate vulnerability identification, actionable fault localization, and verified loop optimization.