Frequency-based trigger injection is a technique in machine learning security where backdoor triggers are embedded into input data by modifying its frequency domain representation rather than altering its spatial or temporal values directly. By applying mathematical transformations such as the discrete cosine transform or Fourier transform, an attacker subtly alters specific spectral components to implant a hidden trigger pattern into media like images or audio. This approach preserves the semantic integrity and perceptual appearance of the poisoned data, making the embedded attack stealthy against standard visual inspections and spatial-domain defense filters while ensuring the compromised model reliably executes unintended, malicious behaviors whenever the targeted frequency pattern is encountered.
