An improved genetic algorithm and its application in neural network adversarial attack.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning Security

Background:

• Genetic algorithms (GAs) are widely used optimization tools.
• Crossover and mutation strategies significantly impact GA performance.
• Existing GAs may face challenges in searchability, convergence, and precision.

Purpose of the Study:

• To propose a novel improved genetic algorithm (GA) with enhanced crossover and mutation operations.
• To evaluate the proposed GA's performance against mainstream swarm intelligence algorithms.
• To assess the algorithm's applicability in generating neural network adversarial attacks.

Main Methods:

• Modification of crossover and mutation operators in a simple genetic algorithm.
• Empirical validation using 15 diverse test functions.
• Comparative analysis with three other swarm intelligence optimization algorithms.
• Application to neural network adversarial attack generation using only output information.

Main Results:

• The improved GA exhibited enhanced global search ability, convergence efficiency, and precision.
• Qualitative results showed superiority over three mainstream swarm intelligence algorithms.
• Quantitative results indicated superior performance on 13 out of 15 test functions.
• Statistical evaluation using Wilcoxon rank-sum test confirmed significant advantages at 95% confidence intervals.
• Successful generation of adversarial samples for neural networks with high confidence and speed, without requiring internal model information.

Conclusions:

• The proposed improved genetic algorithm offers significant advancements in optimization capabilities.
• The enhanced GA provides a robust and efficient method for tackling complex optimization problems.
• The algorithm is effective for generating neural network adversarial attacks, demonstrating its versatility and practical utility.