The proliferation of AI-generated manuscripts, paper mill publications, and fabricated experimental data has precipitated a crisis of integrity in academic publishing, with more than 50,000 retracted papers catalogued by the Retraction Watch database to date. Existing detection tools, including GPTZero and the Turnitin AI detection module, operate exclusively on surface-level textual features, rendering them vulnerable to sophisticated paraphrasing attacks and categorically blind to structural manipulation in citation networks. This paper introduces FakeScholar, a multimodal deep learning framework that jointly analyses paper text content, citation graph topology, author collaboration networks, and reported statistical anomalies to classify academic papers as genuine or fraudulent. FakeScholar integrates a fine-tuned SciBERT encoder, a three-layer Graph Convolutional Network (GCN) operating over citation graphs, a GraphSAGE-based author network encoder, and an Isolation Forest statistical anomaly detector. The four modality representations are fused via a two-layer multilayer perceptron (MLP) to produce a calibrated binary fraud probability. Evaluated on a curated benchmark dataset of 50,000 papers equally balanced between genuine and fraudulent samples drawn from Retraction Watch, known paper mill repositories, synthetically generated AI papers, arXiv, PubMed Central, and the Semantic Scholar Open Research Corpus, FakeScholar achieves an F1 score of 0.912, an AUROC of 0.961, and an accuracy of 91.4%, outperforming the strongest single-modality baseline by 12.3 percentage points in F1. Ablation studies confirm that the citation graph encoder contributes the largest individual performance gain, its removal causing a 10.7 percentage-point decline in F1. Robustness experiments demonstrate that FakeScholar maintains an F1 of 0.874 on aggressively paraphrased AI-generated papers, compared to 0.631 for text-only baselines. FakeScholar introduces a unified multimodal architecture for identifying academic misconduct, providing the groundwork for seamless real-time integration into scholarly publishing and manuscript evaluation platforms.

Index Terms—academic fraud detection, multimodal deep learning, citation graph analysis, AI-generated text detection, SciBERT, graph neural networks, paper mills