Despite the rapid proliferation of image-generating Generative AI (GenAI) architectures, research into backdoor attacks and defences of their underlying diffusion models remains limited. Although previous work has investigated detecting and removing backdoor triggers in the input noise, these defences have assumed that triggers will occur well outside the normal noise distribution. In this work, we not only build upon existing trigger embedding strategies with our EMPDiffusion approach, but we also propose our Fourier transform-based SpecDet as a method to detect potentially malicious noise across a wide variety of attack techniques. We further leverage detected malicious noise to reconstruct the original trigger, and we remove backdoor behaviour entirely via our Lost In The Noise (LITeN) retraining pipeline. We conduct a variety of empirical experiments to show that EMPDiffusion bypasses existing detection and reconstruction techniques, while SpecDec and LITeN successfully reduce the attack success rate (ASR) to 0% on both our EMPDiffusion and other existing backdoor trigger techniques.