Is frequency-based detectors superseded?

frequency-based detectors (Parameter-efficient fine-tuning (LoRA family)): cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here. 2 paper(s) critique it, 0 beat it on benchmarks — #234 of 1113 most-superseded. Sub-problem: cluster led by CNN. Newer alternatives in the same sub-problem include Structured Convolutional Projection + LoRA, LEGO, layer-selective multimodal large language models (MLLMs) with contrastive LoRA tuning and layer sensitivity analysis (LSA), AdaptPrompt.

Parameter-efficient fine-tuning (LoRA family)

frequency-based detectors

Superseded baseline#234 of 1,113 most-superseded

Cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here

2 papers critique it · 0 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites frequency-based detectors as a baseline.

While these methods proved highly effective in controlled environments, they exhibited severe fragility when exposed to standard image processing operations. For instance, a simple JPEG compression pass acts as a low-pass filter, effectively erasing the high-frequency clues that frequency-based detectors depend upon

— Boosting Robust AIGI Detection with LoRA-based Pairwise Training

“Unlike GANs, diffusion models generate images via iterative denoising, resulting in fingerprints that resemble high-frequency Gaussian noise rather than structural periodicity. This shift renders frequency-based detectors less effective on modern deepfakes (e.g., Stable Diffusion rombach2022latentdiffusion, DALL-E openai2023dalle3)”

— AdaptPrompt: Parameter-Efficient Adaptation of VLMs for Generalizable Deepfake Detection

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.