Abstarct: With the increasing complexity of modern software systems, automated detection of security vulnerabilities in source code has become a critical challenge in cybersecurity. While recent approaches baxsed on Deep Learning and Graph Neural Networks (GNNs) have shown promise, modeling code using standard Code Property Graphs (CPGs) often results in excessively large and noisy graphs. This structural complexity leads to issues such as "state space explosion" and "oversmoothing," thereby diminishing the model's ability to distinguish subtle vulnerability patterns. To address these limitations, this thesis proposes a novel frxamework named HyperSlice, which synergizes Program Slicing with Hypergraph Neural Networks (HGNNs). In this approach, semantic slices are first extracted baxsed on sensitive program points. Subsequently, instead of decomposing relationships into numerous binary edges, each slice is encapsulated as a single "hyperedge" within a hypergraph structure. This compact representation preserves integral data and control dependencies while reducing the number of structural relations by 96.7% compared to full CPGs. Semantic features are extracted using the pre-trained CodeBERT model. Experimental evaluation on the standard Devign dataset (comprising FFmpeg and QEMU open-source projects) demonstrates that HyperSlice achieves an F1-score of 61.47%, delivering performance competitive with state-of-the-art baxselines. A significant achievement of this research is the substantial improvement in Recall, reaching 77.99%, which marks an approximate 6.8% increase over the Devign baxseline. This result highlights the proposed method's efficacy in mitigating false negative rates and accurately identifying hidden vulnerability patterns, which is of paramount importance in security-critical applications.