Backdoor attacks poison a model's training data so that the model behaves normally on clean inputs but switches to an attacker-chosen label when a specific trigger appears. A striking recent family of attacks uses the trigger of simply rotating the input image: Wu et al. showed that poisoning about four images (0.01% of GTSRB) rotated by 45° yields an attack success rate of nearly 70% with under 1% clean accuracy loss, and Xu et al. pushed this to ASR often above 99.5% by pairing a fixed trigger angle with randomly rotated benign data. Both papers close by calling for "truly transformation-invariant architectures" as a defense. Separately, a line of work on geometric invariance has built networks that are mathematically rotation-invariant: they map the image to a polar-frequency representation and discard the angular phase, so a rotation by any angle leaves the features bit-for-bit unchanged in the continuous limit (Delchevalerie et al., Wu et al., Ryu et al.). Nobody has ever run a backdoor attack on such a network.
This proposal tests a claim that goes beyond "invariance is a defense": hard-coded invariance neutralizes angle-specific rotation backdoors but does not remove the attack surface — it relocates it into the invariant component of the representation, where existing attacks and defenses have never looked. Concretely, an invariant network provably cannot bind a label to a 45° orientation, so BATT-R and "Just Rotate It" should collapse — but a trigger designed to live entirely in the rotation-invariant radial spectrum should succeed exactly where those attacks fail, and should survive test-time rotation jitter that destroys all Cartesian frequency triggers. The proposal is a paired study: a negative result (transformation backdoors die under exact invariance) and a positive result (a new attack class, the radial spectral trigger, that exploits the invariance itself).
Core Claim
When a network computes features that are invariant to a transformation group, any backdoor whose trigger is an element of that group is undeployable: the network cannot distinguish $\theta^\*$ from $0^\circ$, so no poisoning budget can teach the association. But the invariant features are still a rich code — the radial power profile retains enough shape information to reach 90.20% average precision on airplane detection in satellite imagery (Wu et al.), and DFT magnitude pooling improved ImageNet top-1 error by 0.78–1.81% over standard pooling across AlexNet, VGG-VD 16, Inception-v3, and ResNet (Ryu et al.). Anything the network can represent, it can be poisoned on. The question is whether an attacker can craft an imperceptible perturbation that (i) survives the invariance-inducing projection and (ii) is invisible to defenses. We claim the answer is yes: perturb the image's polar Fourier spectrum along the radial axis only.
Writing the invariant projection as $P(r) = \int_0^{2\pi} |F(r,\theta)|^2 \, d\theta$, where $F(r,\theta)$ is the image's spectrum in polar frequency coordinates, a radial trigger is a perturbation $\delta(r)$ added to the spectrum before pooling, constructed to depend only on radius $r$ and not angle $\theta$. Because a spatial rotation traces circles at constant $r$, this trigger is exactly rotation-invariant in the continuous case — the same mathematical fact that makes the architecture invariant now makes the attack invariant. This is the mirror image of Gap 3 in the literature: existing frequency attacks inject into specific Cartesian bins $(u,v)$ (Wang et al., FTrojan; Zeng et al.), which scatter across bins when the input rotates. Nobody has built a trigger in the radially pooled, angle-free subspace.
Research Program
The program has three parts. First, the immunity audit: we implement two mathematically rotation-invariant backbones — a Bessel-convolutional network following Delchevalerie et al. and a polar-Fourier magnitude network in the style of Wu et al. — plus a ResNet with DFT magnitude pooling (Ryu et al.), and run BATT-R and "Just Rotate It" poisoning against them at matched poisoning rates. Our prediction is near-chance ASR with intact clean accuracy, but the honest open question is discretization: Delchevalerie et al. document aliasing and interpolation errors on discrete grids, so small residual angle sensitivity might let a rotation backdoor partially survive through the aliasing back door. Measuring how much ASR leaks through discretization error, as a function of image resolution, is itself a novel quantitative result (it connects to Gap 5, since Zeng et al. found frequency effects are strongly resolution-dependent).
Second, the radial spectral trigger. We craft $\delta(r)$ by bilevel optimization in the spirit of Zeng et al., who used bilevel optimization to build smooth low-frequency triggers achieving ~97% ASR on CIFAR-10 while evading frequency detectors; our variant constrains the perturbation to be constant along $\theta$ and concentrates its energy at radii where natural images carry little power, so it is both stealthy and rotation-proof. We evaluate it against invariant and standard networks alike: on a standard CNN it competes with FTrojan (98.78% average ASR, 0.56% clean accuracy drop, Wang et al.); on invariant networks it should be the only spectral attack that works, because FTrojan's Cartesian-bin perturbations are annihilated by angular pooling.
Third, the defense implications. We test a cheap countermeasure the literature has never evaluated: random rotation jitter at inference. Because FTrojan-style triggers live in Cartesian bins that rotate with the image, jitter should destroy them — a defense costing nothing but a resampling step. The radial trigger is the control that survives it. We also stress-test the strongest detectors against both trigger types, and probe UNICORN specifically: its inversion framework learns an invertible map $\phi$ and its inverse $\phi^{-1}$, but phase discarding and radial pooling are many-to-one — $\phi^{-1}$ does not exist mathematically — so trigger inversion should have a principled blind spot here.
Baselines
The do-nothing reference point is the standard CNN under transformation backdoors: Wu et al. report nearly 70% ASR on GTSRB at 45° from 0.01% poisoning (about four images) with under 1% clean accuracy drop, and Xu et al. report ASR often above 99.5% with benign accuracy typically above 90% on ResNet-18 over CIFAR-10 and GTSRB, using a 16° rotation trigger with randomly rotated benign training data. These define the vulnerability level that invariance is supposed to remove; beating them means driving ASR to chance while matching their clean accuracy at the same poisoning rate, data, and tuning budget.
The strongest existing attack competitor is BATT-R itself (Xu et al.), because it was explicitly engineered to defeat rotation augmentation and patch-synthesis defenses; we run its exact poisoning protocol (fixed trigger angle plus uniformly rotated benign samples) against each invariant backbone under identical data splits and poisoning fractions. For the radial trigger, the strongest attack competitor is FTrojan (Wang et al.), matched in perceptual magnitude budget and poisoning rate; beating it means matching its ~98.78% ASR on invariant networks and retaining that ASR under test-time rotation, where its Cartesian trigger should collapse.
The strongest defense competitor is the frequency-perspective detector of Zeng et al., which reports an average backdoor detection rate of 98.5%; because it found detection works far better at high resolution, we evaluate it at its best setting (PubFig at 224×224) rather than only on CIFAR-10's 32×32 images. We also run Fine-Pruning (Liu et al.), which reduced ASR from 100% to 0% on YouTube Aligned Face with negligible accuracy loss, since dormant-neuron pruning is the natural repair once a backdoor is found. The key ablation isolates radial symmetry as the ingredient: the same perturbation energy, the same frequency band, allocated either isotropically in $r$ or into fixed Cartesian bins. If only the radial allocation survives rotation and angular pooling, the mechanism — not the magnitude or band — is what matters.
Why not just…?
Why not just use random rotation augmentation instead of invariant architectures? The file's evidence says augmentation is unreliable and sometimes harmful. Wu et al. found augmentation is a double-edged sword: trigger angles outside the augmentation range become more vulnerable, and aggressive rotation augmentation hurts tasks where orientation is semantic, like telling left-turn from right-turn signs. More decisively, Xu et al. built BATT specifically to defeat augmentation by rotating benign data randomly while keeping the trigger angle fixed, and still got above 99.5% ASR. Exact invariance sidesteps all of this because it removes orientation from the representation entirely rather than diluting it with data. That said, augmentation remains the cheap baseline every reviewer will demand, so we run it as a comparison arm.
Why not just use existing run-time defenses (STRIP, Neural Cleanse, frequency detectors, Fine-Pruning)? Each fails on a documented axis for this threat model. STRIP (Gao et al.) was bypassed by FTrojan's dispersed patterns (Wang et al.) and by representation-aligned trigger families (Lilith, Zhou Feng et al.); Neural Cleanse assumes a localized patch and misses global transformations (Wu et al., Xu et al.). The real threat is Zeng's frequency detector with its 98.5% detection rate, and we do not argue around it — we test it head-on at its strongest resolution setting. It may well catch our radial trigger; if it does, we will apply Zeng's own lesson that bilevel-optimized smooth triggers evade their detector (they achieved ~97% ASR while evading it) and report honestly whether the radial constraint changes that. Fine-Pruning is complementary, not competing: it repairs a model after detection, and we include it for completeness.
Why not just rotate the input at test time to break all frequency triggers? This is a genuinely strong simple defense, and part of our contribution is to establish its reach and its limit. It should neutralize every Cartesian-bin trigger, including FTrojan, at near-zero cost — a useful finding if confirmed. It cannot touch the radial trigger, which is rotation-invariant by construction, and that contrast is precisely the experiment that proves the invariant subspace is a real attack surface rather than a curiosity. So we run it as a first-class defense arm rather than a footnote.
Why not just assume UNICORN's trigger inversion already covers non-patch backdoors? UNICORN (UNICORN) reports an average inverted-trigger ASR of 95.60% and stays above 90% detection where Neural Cleanse and TABOR fall below 60% on warping attacks. But its framework learns a map and its inverse; angular pooling and phase discarding are mathematically non-invertible. We will run UNICORN against the radial trigger and expect a principled failure, with its known 2.7× slowdown over Neural Cleanse making the practical cost of that failure worse.
Proposed Experiments
Datasets and architectures follow the file's benchmark map: GTSRB and CIFAR-10 as the common ground of both subfields; PubFig at 224×224 as the high-resolution regime where Zeng et al. say frequency detection is strongest; Rotated MNIST and Outex-TC-00010-r for the invariant backbones, where Delchevalerie et al. report B-CNNs beating augmented group-CNNs by nearly 15% at 25% training data; and a physical-world check photographing real traffic signs at arbitrary orientations, mirroring the physical validations in Xu et al. and Wu et al.. Metrics are attack success rate, benign accuracy, detector backdoor detection rate, and ASR under test-time rotation sweeps of the trigger. Three checks are decisive. First, BATT-R and Just-Rotate-It poisoning on the invariant backbones at the published poisoning rates: if ASR stays near chance with intact clean accuracy, the negative half of the claim is established, with the discretization-leakage measurement quantifying exactly how much invariance the discrete grid really delivers. Second, the radial-versus-Cartesian ablation on the same networks under test-time rotation: only the radial trigger should hold ASR across the full rotation sweep. Third, the defense battery — rotation jitter, Zeng's detector, STRIP, Fine-Pruning, and UNICORN — run against both trigger types, with UNICORN's non-invertibility blind spot tested as a falsifiable prediction rather than an assumption. If the frequency detector catches the radial trigger even after bilevel smoothing, the proposal still delivers its first two findings, and the defense picture becomes the headline: invariance kills transformation backdoors, and known detectors kill spectral ones, leaving the map of what remains as the contribution.
References
- Just Rotate it: Deploying Backdoor Attacks via Rotation Transformation
- BATT: Backdoor Attack with Transformation-Based Triggers
- Achieving Rotational Invariance with Bessel-Convolutional Neural Networks
- Fourier-based Rotation-invariant Feature Boosting: An Efficient Framework for Geospatial Object Detection
- DFT-based Transformation Invariant Pooling Layer for Visual Classification
- Backdoor Attack through Frequency Domain
- Rethinking the Backdoor Attacks' Triggers: A Frequency Perspective
- UNICORN: A Unified Backdoor Trigger Inversion Framework
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- STRIP: a defence against trojan attacks on deep neural networks
- Lilith: Backdoor Generalization under Training-Inference Trigger Shift