Backdoor attacks poison a small fraction of a training set so that a model behaves normally on clean images but flips to an attacker-chosen label whenever a secret trigger is present. The hardest versions of these attacks no longer paste a visible patch onto the image; they perturb the image's frequency representation — the coefficients of a Fourier or cosine transform — so that a tiny change in a few spectral coefficients smears an imperceptible ripple across every pixel. Existing defenses against these attacks detect poisoned samples by looking at where the model is sensitive in frequency space: Don't FREAK Out shows that if you compute, per sample, how the model's loss responds to changes in each Fourier magnitude coefficient, then poisoned samples separate from clean ones — but it separates them with Principal Component Analysis, which only finds directions of maximum variance and therefore collapses exactly where it matters most. Low-frequency attacks such as FIBA hide the trigger inside the low frequencies that models legitimately use for semantics, and the variance-based separation breaks down.
This proposal imports a mature statistical machinery to that breakdown point. Independent Component Analysis (ICA) — the technique that separates overlapping voices in a noisy room, or overlapping brain networks in an fMRI scan — separates signals that are statistically independent and non-Gaussian, not merely uncorrelated. The core claim is that a frequency-domain backdoor trigger, viewed correctly, is an independent component: a fixed, sparse, strongly non-Gaussian spectral pattern that is mixed into the benign spectral statistics of a subset of training samples with an unknown scaling. Neuroimaging has spent two decades solving precisely this shape of problem — separating a repeated spatial pattern from a large population of noisy mixtures, with known indeterminacies in ordering and scale, solved by constrained and template-guided ICA. Nobody has applied it to backdoor defense, and the fit is unusually tight.
Core Claim
The claim has three parts. First, detection: if we stack the per-sample frequency-sensitivity maps of a training population into a matrix and run group ICA (the multi-subject procedure from neuroimaging), the trigger emerges as a single independent component whose per-sample loading cleanly flags poisoned samples — including for low-frequency amplitude attacks where PCA-based detection degrades. Second, quantification: because constrained ICA resolves the scale indeterminacy of standard ICA, the estimated loading is an estimate of the attacker's injection strength — the blend ratio $\alpha$ in FIBA-style amplitude injection or the perturbation intensity (0.06 high-frequency vs. 0.8 low-frequency) measured by Rao et al. — which no current defense recovers. Third, purification: rather than zeroing out suspicious frequencies and inpainting (which fails at ImageNet resolution; Don't FREAK Out: A Frequency-Inspired Approach to Detecting Backdoor Poisoned Samples in DNNs), we subtract the estimated trigger component from each flagged sample's spectrum — a soft demixing that removes the attacker's signal while preserving the benign spectral content, including the fine high-frequency detail that hard masking destroys.
The proposal differs from the nearest existing work in its statistics and in what it produces. FREAK uses second-order PCA plus a Wasserstein distance for detection and binary masking plus a reconstruction network for purification; it detects and discards, but never isolates the trigger itself. ASSET separates clean from poisoned samples by inducing loss divergence but cannot unmix or reconstruct the trigger. RFLBAT uses PCA-plus-clustering on federated gradients, again second-order. The frequency-domain ICA literature itself — permutation alignment for frequency-domain ICA and time-frequency masking for dominant-source extraction — solves unmixing in acoustics but has never touched adversarial learning. Our proposal is, to our knowledge, the first to treat the poisoned training set as a blind source separation problem in which the trigger is an estimable source signal with a recoverable waveform and a recoverable per-sample amplitude.
Research Program
The mixture model. The key modeling insight is that the mixing really is linear, which is what makes ICA applicable rather than aspirational. FIBA blends a fixed trigger amplitude spectrum into the benign amplitude spectrum with a linear blend ratio $\alpha$; FTrojan adds a fixed pattern to specific DCT coefficients. Both are additive linear mixtures in coefficient space — the exact regime $\mathbf{x} = \mathbf{A}\mathbf{s}$ that ICA was built for. We build a population-level data matrix in the spirit of fMRI analysis, where a scan is a matrix of voxels by time points decomposed into independent spatial maps with associated time courses:
$$\mathbf{X} \approx \hat{\mathbf{S}}\,\hat{\mathbf{M}}^T$$
Here $\mathbf{X} \in \mathbb{R}^{F \times N}$ stacks $F$ frequency coordinates (the analog of voxels) against $N$ training samples (the analog of time points); the columns of $\hat{\mathbf{S}}$ are independent spectral components, and row $i$ of $\hat{\mathbf{M}}$ gives sample $i$'s loading on each component. The trigger should appear as one column of $\hat{\mathbf{S}}$: a sparse, fixed spectral pattern with near-zero loading on clean samples and a consistent loading on poisoned ones — structurally the same object as an intrinsic connectivity network in spatial ICA of fMRI, detected from spontaneous fluctuations without any seed or prior. Two signal channels will be compared: raw Fourier amplitude spectra, and the per-sample Fourier-magnitude loss-gradient sensitivity maps that FREAK introduced (where the trigger's footprint is amplified by the model's own memorized sensitivity).
Template guidance for the low-frequency overlap. The reason PCA fails on FIBA is that the trigger's low-frequency variance mixes into the same directions as benign semantic variance. Higher-order statistics help — a fixed blended amplitude pattern is far more sparse and heavy-tailed than natural spectral variation — but we do not rely on that alone. We borrow template guidance from spatially constrained ICA in neuroimaging, where a reference template (NeuroMark) anchors component extraction so that known networks are absorbed by the template and only genuinely novel structure is left to the unmixing; this is what makes single-subject decomposition reliable even from three minutes of scan data. Concretely: from a small clean validation set we estimate the benign spectral components (or their class-conditional templates), constrain them in the ICA objective following the augmented-Lagrangian formulation of Lu & Rajapakse, and let the remaining degrees of freedom capture whatever independent, non-Gaussian, sample-correlated pattern survives — the trigger candidate. Unit-norm equality constraints on the demixing rows fix the scaling so that loadings are comparable across samples (the same mechanism that locked output variances to 1.00 and improved recovered-signal SNR from 21.58 dB to 25.04 dB in CICA); kurtosis-based inequality constraints order components so sparse, high-kurtosis spectral patterns — the signature of concentrated frequency injections — surface first (CICA ordered components by kurtosis from 32.82 down to −1.27 in synthetic mixtures). Where benign and trigger components are partially correlated, sparse BSS variants designed for that regime, such as SparseICA and sparse BSS for partially correlated sources, replace vanilla FastICA.
Detection, quantification, and purification. Detection is a threshold on the trigger component's per-sample loading (with the threshold calibrated on the clean validation set, mirroring how FREAK calibrates against a clean baseline). Quantification is the loading value itself: the recovered estimate of the injection strength. Purification is soft spectral subtraction on flagged samples. If $\hat{\mathbf{s}}_{\text{trig}}$ is the estimated trigger component and $m_i$ sample $i$'s loading, the purified amplitude spectrum is
$$\mathcal{A}'_i = \mathcal{A}_i - m_i\,\hat{\mathbf{s}}_{\text{trig}}$$
This removes only the attacker's contribution and leaves every benign coefficient — including high-frequency detail — untouched, which is precisely what the masking-plus-R2U-Net approach of FREAK cannot do at high resolution. Because we generate all poisons ourselves, we can verify the unmixing directly: the ground-truth trigger spectrum is known, so we measure cosine similarity between $\hat{\mathbf{s}}_{\text{trig}}$ and the true trigger, and the error of $m_i$ against the true injection strength — checks of the BSS claim itself, not just of downstream detection.
Baselines
The do-nothing reference point is the undefended poisoned model. We reproduce each attack at its published strength and report the attack success rate (ASR) and clean/benign accuracy as the starting line: for FTrojan, an average ASR of 98.78% across its five-benchmark suite with an average benign-accuracy drop of only 0.56%; for FIBA, 99.53% ASR on ISIC-2019 and 71.44% on KiTS-19, surviving Fine-Pruning with ASR above 90% even after 80% of neurons are pruned; for the LFC semantic attack, 100% ASR across ResNet18, VGG16, and MobileNetV2. These numbers define what "undefended" means and confirm our attack reproductions are faithful.
The strongest existing competitor is FREAK, the only prior defense that is both frequency-aware and validated against frequency-domain attacks. It reports a true positive rate of 99–100% at a false positive rate below 2% against CYO and FTrojan on ImageNet with ResNet18/34, and 96.6% TPR against BadNet; against FIBA it documents a narrowed sensitivity gap and overlapping distributions, with no reported result found on Lacuna for FIBA detection. We will run FREAK on exactly our poisoned models, datasets, and tuning budget (its top-$k$ frequency selection and Wasserstein threshold swept over the same grid), so the comparison isolates the statistical method, not implementation effort. Beating FREAK means, concretely, higher detection AUROC against FIBA at matched false positive rates, and post-purification clean accuracy at ImageNet resolution where FREAK's masking-plus-reconstruction pipeline is reported to struggle — for which no reported result exists on Lacuna beyond the qualitative failure statement.
Three further baselines complete the picture. Low-pass filtering and JPEG compression, evaluated by Hammoud & Ghanem against frequency attacks, test whether crude spectral excision suffices; they should fail on FIBA, which lives in the low frequencies. PGBD — a post-training defense that penalizes activation drift toward class prototypes, reporting 93% to 99.9% ASR reduction on BadNet and WaNet with about a 7% clean-accuracy drop — will be run on our frequency-poisoned models to test whether model-level sanitization transfers to spectral triggers it was never evaluated on. Finally, the decisive ablation: replace ICA with PCA on the identical population matrix and pipeline, everything else held fixed. This isolates whether higher-order independence is what buys detection on FIBA. If PCA does as well on FIBA in our pipeline, the core claim is wrong and we will say so.
Why not just…?
Why not just use FREAK's PCA pipeline? Because its own authors document where it breaks: against low-frequency amplitude injection (FIBA), the sensitivity distributions of clean and poisoned samples overlap and detection degrades, with no reported number found on Lacuna for FIBA — precisely the regime our template-guided, higher-order unmixing targets. And its purification by binary frequency masking plus R2U-Net reconstruction succeeds on CIFAR-10 but fails to scale to high-resolution ImageNet images, destroying high-frequency detail. Our pipeline replaces both halves: independence instead of variance for detection, subtraction instead of masking for purification. The honest caveat is that FREAK's failure could stem from its per-sample Wasserstein-threshold design rather than from second-order statistics per se — our PCA ablation on the same population matrix is designed to settle exactly that question, and we treat it as a real possibility rather than arguing around it.
Why not just low-pass filter or JPEG-compress the inputs? This works only against the high-frequency attacks. Hammoud & Ghanem evaluated JPEG compression and augmentation-based defenses directly and found that adaptive adversaries bypass them; FIBA and the evolutionary low-frequency attack place the trigger in the low band that any semantics-preserving filter must keep, and Rao et al. show low-frequency injection simply needs a much larger intensity (0.8 vs. 0.06) rather than being impossible. Filtering either misses the trigger or destroys the signal it is riding on. We will run this baseline anyway, because it is cheap and it quantifies the trade-off curve our soft subtraction is meant to dominate.
Why not just use the standard spatial defenses? The file is unambiguous that they are already defeated by the attacks we target: FTrojan bypasses Neural Cleanse, STRIP, Februus, and NAD because its anomaly index stays below Neural Cleanse's threshold and its perturbation is globally dispersed rather than spatially localized; FIBA bypasses Neural Cleanse and STRIP and keeps ASR above 90% after 80% Fine-Pruning; Rao et al. show Grad-CAM saliency maps for poisoned images are indistinguishable from clean ones, leaving Februus nothing to mask. These defenses fail for a structural reason — they look for localized or spatially anomalous evidence — not a tunable one, so they are not serious competitors here.
Why not just induce loss divergence and threshold, as ASSET does? This is the most credible threat to the detection half of the proposal. ASSET is paradigm-agnostic, works from as few as ten clean samples, and improved average detection by 69.3% in self-supervised and 33.2% in transfer learning settings across 56 attack configurations, including clean-label attacks where Spectral Signatures and Activation Clustering fail completely. But ASSET produces a binary verdict per sample: it cannot isolate, quantify, or subtract the trigger, so it cannot purify — and its scalar loss signal ignores spectral structure entirely, which is where FIBA hides. We will run an ASSET-style offset-optimization baseline under the same budget; if it matches our detection AUROC on FIBA, our contribution narrows honestly to quantification and soft purification, which ASSET cannot do by construction.
Proposed Experiments
The attack suite spans the full frequency landscape, all reproduced from the original papers: FTrojan (chrominance DCT injection, high-frequency), the Fourier-mask and additive-filter attacks (sensitivity-targeted, ASR above 90% at 0.1–1% poisoning), the LFC semantic attack (high-frequency semantic injection, 100% ASR), and — the centerpiece — FIBA on ISIC-2019 (99.53% ASR) and KiTS-19 (71.44% ASR) as the low-frequency hard case. Datasets are CIFAR-10, CIFAR-100, GTSRB, an ImageNet subset (the 16-class subset used by FTrojan, plus ImageNet-100 following ASSET), and ISIC-2019; models are ResNet18/34, with VGG16 and MobileNetV2 for architecture robustness following Rao et al.
Metrics come in three tiers. Detection: AUROC and TPR at fixed FPR (1% and 5%), reported per attack, with FREAK's published 99–100% TPR at under 2% FPR on CYO/FTrojan as the bar to match on high-frequency attacks and FIBA as the bar to beat (no reported FIBA number exists). Unmixing fidelity, unique to this proposal: cosine similarity between the estimated trigger component and the ground-truth trigger spectrum, and relative error of the estimated loading against the true injection strength. Purification: post-purification ASR of a model retrained on the purified set, clean-accuracy retention, and PSNR/SSIM of purified images against the originals — comparing soft subtraction to FREAK's binary masking at CIFAR and ImageNet resolutions, where its approach is reported to fail.
Two adaptive attacks close the evaluation, because a defense that only works against the published trigger recipes is not a defense. First, a Gaussianized trigger: ICA's identifiability rests on at most one Gaussian source, so an attacker who randomizes trigger coefficients toward a Gaussian distribution attacks the core assumption head-on; here the template constraint must carry the defense, and if it cannot, that is a genuine and reportable boundary. Second, a distributed trigger: an attacker who spreads the injection across many weak components attacks the sparsity prior; kurtosis ordering constraints and multi-component loadings are the response. Both are cheap to run against our own pipeline, and we commit in advance to reporting the regime where higher-order statistics stop helping — the PCA ablation, the Gaussianized trigger, and the ASSET comparison together determine whether the idea survives contact with a serious adversary.
References
- Don't FREAK Out: A Frequency-Inspired Approach to Detecting Backdoor Poisoned Samples in DNNs
- FIBA: Frequency-Injection based Backdoor Attack in Medical Image Analysis
- Constrained Independent Component Analysis
- AS-FIBA: Adaptive Selective Frequency-Injection for Backdoor Attack on Deep Face Restoration
- Rethinking CNN’s Generalization to Backdoor Attack from Frequency Domain
- ASSET: Robust Backdoor Data Detection Across a Multiplicity of Deep Learning Paradigms
- RFLBAT: A Robust Federated Learning Algorithm against Backdoor Attack
- Permutation Alignment for Frequency Domain ICA Using Subspace Beamforming Methods
- Blind Extraction of Dominant Target Sources Using ICA and Time-Frequency Masking
- An Invisible Black-Box Backdoor Attack Through Frequency Domain
- Alterations in Memory Networks in Mild Cognitive Impairment and Alzheimer's Disease: An Independent Component Analysis
- Reliability and clinical utility of spatially constrained estimates of intrinsic functional networks from very short <scp>fMRI</scp> scans
- SparseICA: Sparse Independent Component Analysis
- Sparse blind source separation for partially correlated sources
- Check Your Other Door! Creating Backdoor Attacks in the Frequency Domain
- Prototype Guided Backdoor Defense
- Low-Frequency Black-Box Backdoor Attack via Evolutionary Algorithm