The determination of true authorship is critical in the fields of digital forensics, preserving academic honesty, and analyzing historical texts. Such tools of verifying authorship as hand-crafted features are quite inefficient when facing problems of adversarial mimicking text fragmentation and domain shifts. This study offers a robust, scalable framework that combines Siamese neural networks with BERT embeddings, enhanced with CNN-BiLSTM architectures. In addition, it leverages the Impostor Projection Methodology for adversarial training, while utilizing Dynamic Time Warping (DTW), anomaly detection with Isolation Forest and K-Medoids clustering for stylistic, semantic and mimicry variability. This approach addresses many challenges that traditional methods imposed or failed to manage, indicating great potential for authorship verification across numerous domains.
[+] In the PhaseA directory you will find the research and planning documents for the project, along with the presentation slides of the project proposal.