Online exam cheating detection and blockchain trusted deposit based on YOLOv12

In recent years, online examinations have been widely adopted because of their flexibility, but the covert and diverse nature of cheating behaviour poses a serious challenge to the fairness and integrity of examinations. Existing anti-cheating techniques are deficient in detecting diverse cheating behaviours in real-time and ensuring the credibility of evidence. To address this problem, this paper proposes an integrated solution for online exam cheating detection based on the lightweight YOLOv12 model and blockchain trusted depository. Firstly, we made targeted lightweight improvements to the benchmark YOLOv12n model by removing the computationally intensive Attention mechanism from the backbone network and simplifying the module structure (modifying the A2C2f module), as well as replacing the computationally heavy C3k2 module in the head network with the efficient C3Ghost module. These modifications aim to reduce the model's computational complexity and number of parameters, increasing inference speed, thus making it more suitable for real-time detection tasks. Secondly, to address the issue of credible evidence preservation concerning cheating, we constructed a evidence preservation system based on the Hyperledger Fabric consortium blockchain, combined with IPFS distributed storage technology. Key screenshots of suspected cheating behaviors are stored on IPFS, and their content identifier (CID) along with detection metadata (such as timestamp, detection type, confidence, etc.) is recorded on the blockchain through smart contracts, ensuring the originality, integrity, and immutability of the evidence. Experiments conducted on an online exam cheating dataset containing categories of 'person' and 'electronic devices' demonstrate that the proposed lightweight YOLOv12NoAttn model exhibits competitive detection performance (with slight improvements in mAP50 and Recall) while showing higher efficiency by significantly reducing parameters (approximately 28%) and GFLOPs (approximately 13%). Ablation experiments further verify the effectiveness of the lightweight improvements made to both the backbone and head networks. This research provides an efficient, accurate, and trustworthy solution for cheating detection and evidence management in online examinations, contributing to the maintenance of fairness and integrity in online assessments.