Defending Time-Series Models Against Universal Adversarial Attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems fail to effectively defend against universal adversarial attacks on time series data used in machine learning models, which can lead to incorrect assessments and treatments in critical applications like healthcare, due to data corruption.
Innovation Solution
A processor-implemented method and system that identifies data-driven models matching input data regimes, performs universal adversarial attacks by computing and clipping perturbations, and retraining models using data augmentation techniques to maintain performance above a threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data-driven models are used to automate processes and assess data, then productivity and automation are improved, but the systems become vulnerable to adversarial attacks that can corrupt data and lead to incorrect assessments
Solution Approach 1:
The system performs preliminary adversarial attacks during the training phase to generate adversarial examples, which are then used to retrain the model. This preemptive approach strengthens the model's robustness before actual deployment, preventing successful attacks during operational use.
Solution Approach 2:
The system applies adversarial perturbations to training data to create adversarial examples that actively counter potential attacks. By incorporating these adversarial examples into retraining, the model develops resistance against future adversarial attacks, effectively neutralizing the threat before it can exploit the system.
2Reliability
If adversarial attacks are performed to assess model vulnerability, then security assessment is improved, but model performance may deteriorate due to data corruption
Solution Approach 1:
The system converts the harmful adversarial perturbations into beneficial training data. By using adversarial examples as part of the retraining process, the model learns to recognize and resist attack patterns, transforming a security threat into an opportunity for strengthening model robustness and performance.
Solution Approach 2:
The system modifies the training data parameters by incorporating adversarial perturbations with controlled magnitudes (e.g., L-infinity norm constraints). This parameter adjustment allows the model to learn from adversarial examples while maintaining normal operational performance, balancing security assessment with model effectiveness.
3Reliability
If universal adversarial perturbations are computed and applied, then defense against attacks is improved, but the complexity of the system increases due to multiple iterations and clipping operations
Solution Approach 1:
The system performs self-training by automatically generating adversarial examples and using them to retrain the model without requiring external intervention. This self-service approach simplifies deployment and maintenance while achieving robust defense against adversarial attacks through iterative self-improvement.
Solution Approach 2:
The system computes universal adversarial perturbations that can defend against multiple types of attacks simultaneously. This universal approach consolidates multiple defense mechanisms into a single framework, reducing overall system complexity while maintaining comprehensive protection across different attack scenarios.
Data Source
AI summary
Data is prone to various attacks such as cyber-security attacks, in any industry. State of the art systems in the domain of data security fail to identify adversarial attacks in real-time, and this leads to security issues, as well as results in the process/system providing unintended results. The disclosure herein generally relates to data security analysis, and, more particularly, to a method and system for assessing impact of adversarial attacks on time series data and providing defenses against such attacks. The system performs adversarial attacks on a selected data-driven model to determine impact of the adversarial attacks on the selected data model, and if the impact is such that performance of the selected data model is less than a threshold, then the selected data model is retrained.