Authors:

Alisha Verma

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

As artificial intelligence systems are increasingly trained on sensitive personal data, regulations such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA) have established the legal "right to be forgotten," granting individuals the right to request data deletion. Machine unlearning has emerged as a solution to remove specific data influence from trained models without costly full retraining, addressing both individual privacy rights and organizational needs to remove corrupted or biased data that may compromise model performance. However, it remains unclear whether existing methods truly eliminate data influence, particularly in Graph Neural Networks (GNNs), where interconnected structures cause information to propagate across nodes through neighborhood aggregation. This study hypothesizes that current approximate graph unlearning methods fail to completely remove targeted data, leaving residual information detectable through privacy auditing. To test this, we applied multiple state-of-the-art graph unlearning methods—including node and edge unlearning approaches—to benchmark graph. datasets and evaluated each method's completeness using tailored membership inference attacks (MIAs), which attempt to determine whether specific data points were part of the original training set. Our preliminary findings reveal that across all tested methods, MIAs consistently detect traces of supposedly removed data at rates significantly above random chance, indicating persistent information leakage. No approximate method achieved the privacy guarantees of full retraining from scratch. These results suggest that current graph unlearning techniques are insufficient for genuine regulatory compliance and pose ongoing risks to both individual privacy and model integrity, as residual data continues to influence predictions in sensitive domains including healthcare, finance, and social networks. This work contributes to the broader fields of AI governance and trustworthy machine learning by establishing membership inference attacks as a critical auditing framework for evaluating unlearning efficacy and highlighting the urgent need for more robust, verifiable unlearning methodologies. Panel Ill. The Future is Now: Smart Machines, Smarter Solutions 9, Early Detection of Pedicle Screw Loosening Using a Tri-Axial Smart Fixation System Sepehr Khavari,' Yue Wang," Dillan Prasad," John A. Rogers,** and *Christopher S. Ahuja? "Department of Biomedical Engineering, Northwestern University MeCormick School of Engineering, 2145 Sheridan Rd, Evanston, IL 60208 USA "Department of Neurological Surgery, Northwestern University Feinberg School of Medicine, 676 N Saint Clair Street, Chicago, IL 60611 USA email: sepehrkhavari2028@u.northwestern.edu _ *PI: christopher.ahuja@northwestern.edu Spinal fusion is one of the most commonly performed procedures in modern healthcare while continuing to have a high failure rate that can lead to serious neurological complications: when implants fail, spinal stability can worsen and nearby nerves may become compressed, causing pain and functional impairment. During spinal fusion, pedicle screws (threaded metal implants) are placed into the vertebrae to stabilize the spine and are connected across levels using titanium rods. This approach is intended to support proper alignment and promote bone healing. However, repeated mechanical loading over time can cause the screws to loosen, leading to micromotion at the interface between bone and implant. This loosening can accelerate bone damage and increase the risk of nerve injury for individuals. In this study, we demonstrate the feasibility of a tri-axial instability detection system using a benchtop validation of an intelligent pedicle screw design that incorporates inertial measurement units (IMUs) that compare positional deviation between multiple sets of pedicle screws in a vertebral column. The system allows for early identification of screw loosening by continuously tracking small positional changes of the implant along three spatial axes, before clinical symptoms appear. To validate this approach, a smart capacitor model was developed that combines a ring-shaped capacitor where height was determined by the locking force of the pedicle screw with an inductor to create a passive, wireless inductor-capacitor (LC) circuit: changes in screw locking force produced measurable shifts in the circuit's resonant frequency. In addition, bone loss caused by repeated loading showed a strong correlation when measured using both volumetric and gravimetric-based methodologies. These results support the development of smart spinal fixation systems capable of continuous mechanical monitoring, enabling early intervention to help preserve spinal cord and nerve root function.

Source:

University of Illinois Chicago

Topics:

No topics listed

Co-authors:

Alisha Verma