**Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks** **Abstract:** This paper introduces Scissorhands, a novel machine unlearning approach designed to erase the influence of specific data from trained models. Scissorhands identifies the most pertinent parameters in the model relative to the forgetting data using connection sensitivity analysis. These parameters are then reinitialized, creating a trimmed model that minimizes the influence of the forgetting data. The trimmed model is further fine-tuned using a gradient projection-based approach to preserve information on the remaining data while discarding information related to the forgetting data. Experimental results on image classification and image generation tasks demonstrate that Scissorhands outperforms existing methods in terms of performance and efficiency. **Keywords:** Machine unlearning · Connection sensitivity · Diffusion model **Introduction:** The paper addresses the challenge of erasing data influence from trained models in compliance with data regulations like GDPR and CCPA. Scissorhands aims to achieve this by first identifying and reinitializing critical parameters using connection sensitivity, followed by fine-tuning the model to maintain performance on the remaining data. The method is evaluated on various datasets, showing superior performance in forgetting data while preserving model utility. **Methodology:** Scissorhands involves two main phases: trimming and repairing. In the trimming phase, parameters are identified and reinitialized based on their influence on the forgetting data. In the repairing phase, a gradient projection-based approach is used to optimize the model, ensuring that the influence of the forgetting data is minimized while maintaining performance on the remaining data. **Related Work:** The paper reviews existing machine unlearning methods, including retraining from scratch, fine-tuning, and approximate unlearning techniques. It highlights the limitations of these methods and introduces Scissorhands as a more efficient and effective solution. **Experimental Evaluation:** Scissorhands is evaluated on various datasets, including SVHN, CIFAR-10, CIFAR-100, and CelebAMask-HQ, demonstrating its effectiveness in forgetting data while maintaining model performance. The method is also applied to the Stable Diffusion model to eliminate inappropriate content, showing comparable results to state-of-the-art methods. **Conclusion:** Scissorhands is an effective and practical machine unlearning algorithm that balances the objectives of data removal, model utility preservation, and generalization to unseen data. Future work could explore its performance on regression and NLP tasks, as well as its application to sequential data.**Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks** **Abstract:** This paper introduces Scissorhands, a novel machine unlearning approach designed to erase the influence of specific data from trained models. Scissorhands identifies the most pertinent parameters in the model relative to the forgetting data using connection sensitivity analysis. These parameters are then reinitialized, creating a trimmed model that minimizes the influence of the forgetting data. The trimmed model is further fine-tuned using a gradient projection-based approach to preserve information on the remaining data while discarding information related to the forgetting data. Experimental results on image classification and image generation tasks demonstrate that Scissorhands outperforms existing methods in terms of performance and efficiency. **Keywords:** Machine unlearning · Connection sensitivity · Diffusion model **Introduction:** The paper addresses the challenge of erasing data influence from trained models in compliance with data regulations like GDPR and CCPA. Scissorhands aims to achieve this by first identifying and reinitializing critical parameters using connection sensitivity, followed by fine-tuning the model to maintain performance on the remaining data. The method is evaluated on various datasets, showing superior performance in forgetting data while preserving model utility. **Methodology:** Scissorhands involves two main phases: trimming and repairing. In the trimming phase, parameters are identified and reinitialized based on their influence on the forgetting data. In the repairing phase, a gradient projection-based approach is used to optimize the model, ensuring that the influence of the forgetting data is minimized while maintaining performance on the remaining data. **Related Work:** The paper reviews existing machine unlearning methods, including retraining from scratch, fine-tuning, and approximate unlearning techniques. It highlights the limitations of these methods and introduces Scissorhands as a more efficient and effective solution. **Experimental Evaluation:** Scissorhands is evaluated on various datasets, including SVHN, CIFAR-10, CIFAR-100, and CelebAMask-HQ, demonstrating its effectiveness in forgetting data while maintaining model performance. The method is also applied to the Stable Diffusion model to eliminate inappropriate content, showing comparable results to state-of-the-art methods. **Conclusion:** Scissorhands is an effective and practical machine unlearning algorithm that balances the objectives of data removal, model utility preservation, and generalization to unseen data. Future work could explore its performance on regression and NLP tasks, as well as its application to sequential data.