A research team led by Professor Kim Jangho of the Department of Artificial Intelligence at Kookmin University (Cho Hyun-jun, Shin Jong-yun, Lee Do-woon, Kim Jangho) presented “QCU (Quantized Contrastive Unlearning),” a new machine unlearning technology that effectively removes the influence of specific training data while maintaining the performance of AI models, at the 35th ACM International Conference on Information and Knowledge Management (CIKM 2026) held in Rome, Italy, on November 10. This research was accepted as a full research paper and presented as an oral presentation.
​
Machine unlearning is a technology that selectively removes the influence of specific information—such as personal information or sensitive data—from an already trained AI model. Previously, the most reliable approach was to retrain the model from scratch while excluding the relevant data; however, this method has the limitation of requiring a significant amount of time and computational resources. Existing unlearning techniques designed to address this issue often compromised not only the information targeted for deletion but also the model’s general knowledge, creating a trade-off between the effectiveness of data deletion and the preservation of model performance.
​
To resolve these issues, the research team proposed QCU, which combines quantization and contrastive learning. QCU quantizes the intermediate feature maps of an AI model to reduce information specific to individual training data while preserving general semantic information. Based on this, the model is trained to move away from the unique features of the data to be deleted and toward the general knowledge that must be retained. It differs from existing methods in that it specifically guides the model to approach a state closer to that of a model retrained from scratch, excluding the data to be deleted.
​
The research team validated the effectiveness of QCU across various environments, including image classification datasets such as CIFAR-10, CIFAR-100, and ImageNet, as well as large language models (LLMs) and image generation models like Stable Diffusion. Experimental results showed that, compared to existing unlearning techniques, QCU effectively reduced the influence of the information to be deleted while stably maintaining model performance; in experiments evaluating the risk of personal information leakage, it also produced results close to those of a retrained model.
​
This research is significant in that it presents a general-purpose unlearning technique applicable to various fields—such as image recognition, language generation, and image generation—rather than being limited to specific AI models. It is expected to serve as a core technology for ensuring privacy protection and the reliability of AI services in the future.
 
Meanwhile, CIKM is a major international academic conference in the fields of information retrieval, data mining, knowledge management, and artificial intelligence. The title of this paper is “Quantized Contrastive Unlearning,” and the research code is publicly available.