neural architecture search for gans

To address the low efficiency of neural architecture search (NAS) caused by excessively large search spaces, this paper proposes GPT-EA, a collaborative search framework. It introduces generative prior knowledge from large language models (LLMs)—specifically GPT—into NAS for the first time, leveraging GPT to generate semantically coherent and structurally valid network components, thereby substantially shrinking the effective search space. This generative prior is then seamlessly integrated with gradient-free optimization via evolutionary algorithms (EA). The method incorporates key techniques including neural architecture encoding and differentiable/zero-cost proxy evaluations. On multiple benchmarks, GPT-EA outperforms seven hand-crafted architectures and thirteen existing NAS methods. Compared to a GPT-free baseline, fine-tuned GPT-EA achieves up to 12% higher accuracy, demonstrating that generative priors simultaneously enhance both the efficiency and quality of neural architecture search.

To address the low search efficiency of deep generative models in discrete spaces during test-time optimization, this paper proposes Neural Genetic Search (NGS)—a gradient-free, plug-and-play framework. NGS tightly integrates the population-based evolutionary dynamics of genetic algorithms with pre-trained deep generative models (e.g., diffusion and autoregressive models). Crucially, it reformulates crossover as a parent-conditioned generative sampling process—enabling natural compatibility with arbitrary discrete structures. The method comprises four stages: population initialization, selection, generative crossover, and stochastic mutation. Evaluated on three distinct tasks—path planning, adversarial prompt generation for large language models, and molecular design—NGS consistently outperforms existing baselines. Results demonstrate its superior efficiency, robustness to problem formulation, and strong cross-domain generalization capability.

To address the prohibitively high computational cost of training candidate architectures in neural architecture search (NAS), this paper proposes NASGraph—a zero-cost, data-agnostic lightweight method. Its core innovation lies in modeling neural networks as graphs and, for the first time, employing a graph topological feature—average degree—as a performance proxy, thereby eliminating both training and data dependencies. On NAS-Bench-201, NASGraph identifies the optimal architecture among 200 randomly sampled candidates in just 217 seconds with high accuracy. It further achieves competitive ranking correlation across multiple benchmarks, including NAS-Bench-101, NDS, and Micro TransNAS-Bench-101. By decoupling architecture evaluation from training and data, NASGraph significantly improves evaluation efficiency and cross-benchmark transferability, offering a scalable and practical solution for resource-constrained NAS.

Generative Adversarial Network (GAN) generators suffer from excessive model size and lack of general-purpose compression methods for mobile deployment. Method: This paper proposes the first fully automated compression framework tailored specifically for GAN generators. It pioneers the integration of AutoML into GAN compression by jointly optimizing a customized lightweight search space, neural architecture search (NAS), and knowledge distillation—enabling end-to-end discovery of efficient architectures solely from the original generator, without requiring the discriminator or task-specific modifications. The framework is agnostic to GAN architecture and loss function. Results: Evaluated on image-to-image translation and super-resolution tasks, the compressed models achieve several-fold parameter reduction while significantly outperforming existing compression approaches in FID (for generation quality) and PSNR (for reconstruction fidelity), successfully balancing model compactness and perceptual fidelity.

To address the limited effectiveness of DNN testing and retraining caused by image distortion in simulation for safety-critical systems, this paper proposes a closed-loop testing framework integrating the metaheuristic optimizer NSGA-II with a conditional generative adversarial network (cGAN). It is the first work to embed cGAN into a search-driven simulation-based closed-loop testing pipeline, enabling high-fidelity photorealistic image generation, efficient failure-case discovery, and co-adaptive model retraining. The method employs coverage and error sensitivity as test objectives, targeting semantic segmentation DNNs (e.g., DeepLabv3+). Experiments demonstrate a 37% increase in adversarial sample diversity and identify 2.1× more worst-performance triggering images than state-of-the-art methods. After retraining, mean Intersection-over-Union (mIoU) improves by 5.8%, significantly overcoming the efficacy bottleneck of pure simulation-based testing.

Existing meta-neural architecture search (Meta-NAS) approaches suffer from poor generalization, restricted search spaces, and high computational overhead. To address these limitations, we propose GraB-NAS—a novel Meta-NAS framework that pioneers Bayesian optimization-guided graph generation. It models neural architectures as graphs and jointly leverages global Bayesian optimization and local latent-space gradient exploration, thereby eliminating reliance on predefined search spaces and enabling efficient, task-aware architecture discovery. GraB-NAS integrates graph neural networks, variational autoencoders, and meta-learning to support rapid cross-task adaptation. Extensive experiments on multiple benchmarks demonstrate that GraB-NAS significantly outperforms state-of-the-art Meta-NAS methods in both search efficiency and final model performance. Its superior generalization capability and practical effectiveness are empirically validated across diverse tasks and datasets.

To address mode collapse and training instability in generative adversarial networks (GANs), this paper proposes a coevolutionary multi-generator–multi-discriminator framework. It systematically compares evolutionary population update mechanisms—including (μ,λ) and (μ+λ) elitist strategies, as well as tournament selection—in the context of GAN optimization. The key contribution is the identification of the (μ,λ) generational replacement strategy as superior for balancing exploration and exploitation: it mitigates premature convergence induced by elitism while substantially improving sample diversity and support coverage of the target distribution. Experiments on synthetic datasets and MNIST demonstrate that, particularly with large offspring populations, the (μ,λ) strategy achieves significant improvements over baseline methods in both Fréchet Inception Distance (FID) and diversity metrics, alongside enhanced training stability. This work establishes an interpretable, robust coevolutionary paradigm for GAN optimization.

Generative Adversarial Networks (GANs) exhibit a dual role in cybersecurity—both as enablers of adversarial attacks and as promising tools for defense—yet their systematic assessment remains fragmented. Method: This study conducts a PRISMA-compliant systematic literature review (2021–2025.08), analyzing 185 peer-reviewed papers. Contribution/Results: We propose a novel four-dimensional taxonomy encompassing defensive functionality, GAN architecture, security domain, and threat model. Our analysis reveals advances in training stability and task specificity via WGAN-GP, CGAN, and hybrid architectures, and—first among comprehensive reviews—systematically maps GAN-based defenses against LLM-centric threats. Empirical findings indicate that GANs significantly improve accuracy, robustness, and few-shot generalization in intrusion detection, malware analysis, and IoT security. However, critical challenges persist, including training instability, lack of standardized benchmarks, and limited interpretability. This work delivers a structured knowledge graph and actionable research directions for GAN-driven cyber defense.

The rapid advancement of generative AI—including GANs, VAEs, and diffusion models—has led to an overwhelming and fragmented literature, necessitating a systematic synthesis. This survey proposes a unified technical taxonomy that integrates the evolutionary trajectories, architectural variants, and hybridization strategies of these three dominant paradigms, clarifying shared optimization principles for generation quality, diversity, and controllability. It introduces, for the first time, a multi-dimensional classification framework spanning model architecture, training mechanisms, and application domains. Furthermore, incorporating ethical considerations and societal impact, the survey identifies three key frontiers: scalability, trustworthy generation, and human-AI collaboration. By unifying conceptual foundations and highlighting emerging challenges, this work delivers a structured, forward-looking technical roadmap for researchers and practitioners in generative AI.

This work addresses the limited interpretability and controllability of image generation models by proposing an internal mechanism intervention method based on parameterized activation functions. Specifically, we replace standard activations (e.g., ReLU) in mainstream generative architectures—such as StyleGAN2 and BigGAN—with learnable, semantically interpretable parameterized variants (e.g., generalized Swish with shape and bias controls). This enables direct, fine-grained manipulation of activation behavior for targeted image editing, without altering network architecture or requiring additional training. We demonstrate effective, attribute-specific control—including illumination, texture, and pose—on FFHQ and ImageNet. Experimental results confirm that our intervention preserves model fidelity while offering both human-understandable semantics and quantitative effectiveness. The approach establishes a novel paradigm for transparent, plug-and-play control over generative models’ internal representations.