<p>With the rapid advancement of information technology, cybersecurity challenges have become increasingly severe—among which memory vulnerabilities have emerged as a major threat to system stability and security. In particular, heap overflow vulnerabilities are especially problematic in real-world applications due to their high stealth and diverse exploitation techniques. To address this issue, this paper proposes a heap overflow vulnerability detection method based on deep learning and multimodal decision-level fusion. The method targets executable programs by first obtaining binary instruction sequences and assembly code through disassembly. Binary instructions are then mapped into Markov images, from which spatial features are extracted using a transfer learning model. Simultaneously, assembly code is treated as a form of structured text; context-aware word embeddings are generated via a BERT model, followed by deep semantic extraction using a combination of CNN, BiLSTM, and multi-head self-attention mechanisms. To fully leverage the complementary strengths of different modalities, a dynamic gated weighting strategy is employed at the decision level, enabling adaptive allocation of output weights from each modality and dynamic integration based on their respective contributions. Experimental results show that the proposed method achieves an accuracy of 99.49% and an F1-score of 99.36%. It consistently outperforms various mainstream detection techniques in tenfold cross-validation, ablation studies, and comparative evaluations, thereby validating its detection performance and stability.</p>