What is it about?
This study tackles a growing challenge in cybersecurity AI: what happens when someone needs their data removed from a ransomware detection system? Retraining the entire AI model from scratch is slow and costly. This research uses a technique called SISA training, which splits the model's training data into separate chunks ("shards"), so only the affected chunk needs to be retrained when data is deleted. The system uses reinforcement learning agents (DQN and DDQN) trained on real Windows 11 ransomware behavior. Results show the model loses almost no detection accuracy after deletion, and retraining takes a fraction of the usual time.
Featured Image
Photo by Jason Dent on Unsplash
Why is it important?
Privacy laws like GDPR give people the "right to be forgotten," meaning AI systems must be able to erase someone's data on request. Most cybersecurity AI tools ignore this requirement entirely. This research shows it is possible to build ransomware detectors that are both highly accurate and privacy-compliant — without sacrificing speed or security. This matters for hospitals, businesses, and governments that use AI security tools and must also comply with data protection regulations.
Perspectives
Ransomware is one of the most damaging cyber threats facing individuals and organisations today, and I have long been interested in how machine learning can help defend against it. But as AI becomes embedded in critical security infrastructure, I find myself equally concerned about a question that often goes unasked: what happens to people's data once a model is trained on it? This work grew out of that concern. Privacy regulations like GDPR are not just legal formalities — they reflect a genuine ethical obligation to give people control over their own information. I wanted to show that meeting that obligation does not have to mean sacrificing security. The results genuinely surprised me: removing data from a trained ransomware detection model using SISA caused negligible to zero drop in detection accuracy, and retraining only the affected shard took only seconds. I hope this research encourages the cybersecurity community to take responsible AI deployment seriously — not as a compliance checkbox, but as a core design principle. Building systems that are both effective and respectful of privacy is not a trade-off. It is entirely achievable.
Jannatul Ferdous
Charles Sturt University
Read the Original
This page is a summary of: Privacy-Aware Machine Unlearning with SISA for Reinforcement Learning–Based Ransomware Detection, May 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3774905.3794679.
You can read the full text:
Contributors
The following have contributed to this page
