Search for a command to run...
Generative Adversarial Networks are increasingly applied in cybersecurity for generating synthetic attack data, simulating adversarial behaviors, and augmenting anomaly detection models. However, deploying GANs in real-time, resourceconstrained environments such as edge or embedded systems remains challenging due to their computational demands. FPGAs offer a promising platform to accelerate such models due to their low-latency, parallel processing capabilities. In this work, we present a complete and optimized workflow for implementing a trained GAN generator on FPGA using hls4ml and the Xilinx Vivado Design Suite. Our methodology goes beyond standard model conversion by introducing trace-guided quantization, reuse factor tuning, and activation profiling to m aintain numerical fidelity <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$w$</tex> hile <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$m$</tex> inimizing <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$h$</tex> ardware <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$r$</tex> esource <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$u$</tex> sage. <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$T$</tex> he final implementation, synthesized for the Zynq UltraScale+ MPSoC, achieves high efficiency w ith o nly 25 % D SP a nd 8 % LUT utilization, without requiring external memory blocks. While demonstrated on a sine wave dataset for baseline validation, the design flow is directly e xtendable to cybersecurity-relevant tasks, such as system-call sequence generation and adversarial network traffic modeling. This work serves as a hardware-aware foundation for deploying generative models in real-time, edgebased cybersecurity systems.