As cyber-physical systems (CPSs) become an essential part of critical infrastructures and industries, their technological advancement creates a massive space for adversaries. Therefore, it is crucial to sufficiently explore the threat space to assess the systems' resiliency and plan for hardening. Moreover, due to any change in the cyber, physical, or operational level, CPSs often demand a re-analysis of potential threats. Threat analysis by conducting testbed experiments helps comprehend the attack potentiality but is infeasible to explore all attack space. Formal reasoning-based analytics are advantageous for threat analysis, especially for being noninvasive but provable. However, due to the convoluted features and non-linear nature of the system parameters, such formal models also become expensive in solving time, making them unscalable for larger systems. Hence, effective mechanism design is essential to augment the overall performance of the threat synthesis. This work proposes a threat analysis framework, named iAttackGen, where we train generative adversarial networks (GAN) models using the existing stealthy attack dataset (produced from testbed experiments or formal analysis) and generate more attack scenarios. We consider the smart power grid as the reference CPS and stealthy attacks as the threat model. Our evaluation results on standard IEEE bus systems prove iAttackGen's high accuracy and success rate in synthesizing potential threat vectors.