Hugging Face Releases UnAct: A Gradient-Free Class-Unlearning Method

Summary:Hugging Face has introduced UnAct, a novel gradient-free class-unlearning method that removes the influence of specific training data from a model by attenuating the most responsive connections based on forward passes over the images to be forgotten. Unlike existing methods such as SSD and LFSSD, UnAct does not require backpropagation, labels, or retained data, making it more efficient and robust, especially when forget data is scarce. On CIFAR-10 with only five forget images, UnAct's performanc

---

Core Breakthrough

Hugging Face's research team has introduced UnAct, a novel method for efficient machine unlearning. Machine unlearning aims to remove the influence of specific training data from a trained model without retraining from scratch. Existing methods like Selective Synaptic Dampening (SSD) and its label-free variant LFSSD avoid full retraining but still require backpropagation and parameter importance computed over the entire dataset. UnAct addresses these issues by:

• Gradient-Free: Only forward passes over the forget images are needed, eliminating the need for backpropagation.
• No Labels or Retained Data: Relies solely on the responses of the forget images to attenuate the most responsive connections.
• Efficiency: Repeats this process for up to 20 rounds to ensure thorough unlearning.

Technical Highlights

1. Response Scoring Mechanism: UnAct evaluates the responses of late-layer units to identify connections to attenuate.
2. Multi-Round Iteration: Gradually attenuates connections through multiple rounds to ensure complete unlearning.
3. No Gradients, Labels, or Retained Data: Unlike existing methods, UnAct performs well even when forget data is scarce, avoiding network collapse.

Experimental Results

On ResNet-18, UnAct's performance is comparable to SSD and LFSSD on CIFAR-10, CIFAR-20, and CIFAR-100 when forgetting entire classes. However, UnAct never collapses the network when forget data is scarce. In the CIFAR-10 experiment with five forget images, UnAct's distance to retraining is only 0.21 points, while LFSSD and SSD lag behind by 67 and 90 points, respectively.

Industry Impact

UnAct offers a more efficient and reliable solution for machine unlearning, particularly in data-scarce or privacy-sensitive scenarios. Its gradient-free, label-free, and data-free nature makes it ideal for handling sensitive data. Additionally, UnAct's lightweight nature allows for easy integration into existing machine learning workflows, providing a new path for AI system privacy protection and compliance.

Developer Recommendations

• Integrate UnAct into Existing Workflows: Developers can experiment with integrating UnAct into their existing machine learning pipelines to enhance data privacy protection.
• Explore Multimodal Applications: Future research could explore the application of UnAct in multimodal data, such as joint forgetting of images and text.
• Optimize Iteration Rounds: Adjust the number of iteration rounds based on specific application scenarios to find the best balance between unlearning effectiveness and computational cost.

---

Source: Hugging Face Daily Papers (2026-10-03)

— END —

Tags: #Hugging Face Follow #Machine Unlearning Follow #Gradient-Free Follow #Privacy Protection Follow #AI Safety Follow