A backdoored model behaves normally until an input carries a secret "trigger", such as a small patch or a lookalike character. Then it outputs whatever class the attacker chose. Training with differential privacy (DP), which bounds how much any single training example can change the model, is one of the few defenses that reliably stops such triggers from being learned. It also costs a lot: clean accuracy drops, the noise scale must be hand-tuned, and per-example gradient processing can make each epoch up to 80 times slower. Standard DP is symmetric, and that is plausibly where much of the cost comes from. It limits how much one example can raise the model's belief in any class, and equally how much it can lower it. It protects equally against adding and removing an example. A backdoor only needs one of these four: a few added examples that push the probability of one class sharply upward, from roughly its base rate to nearly 1 on triggered inputs.
This note proposes to split DP training into an "upward" channel and a "downward" channel with separate privacy budgets, following the asymmetric-privacy literature that decouples bounds on belief increases from bounds on belief decreases. It then tests whether a strict budget on the upward channel alone suppresses backdoors, while a loose budget on the downward channel recovers clean accuracy. The goal is both a cheaper defense and an answer to a question the backdoor literature has not asked: which component of DP actually does the work against poisoning.
Core Claim
The success of a backdoor is an upward belief shift, or "lift". Lift is the ratio of the model's probability of a class after seeing the trigger to its baseline probability. For a successful all-to-one trigger on a $K$-class problem, lift on the target class is close to $K$. For the source classes it is close to 0. On the Lift, Related Privacy Measures, and Applications to Privacy-Utility Trade-Offs and Asymmetric Local Information Privacy and the Watchdog Mechanism show three things in the data-release setting: worst-case leakage is dominated by upward lift spikes; the upper and lower lift bounds can be budgeted separately; and relaxing the lower bound while keeping the upper one strict buys large utility gains. They report optimal utility when the budget is split with λ = 0.65 toward relaxing the lower bound. That work motivates this proposal but concerns releasing sanitized data, not training models.
Our claim is that the same decoupling carries over to gradient training, where it has a concrete form. The cross-entropy gradient of each example splits exactly into a part that raises the labeled class's score and a part that lowers the other classes' scores. Backdoor implantation should run mainly through the first part. If so, strict clipping and noise on the "raise" part plus loose treatment of the "lower" part should match symmetric DP-SGD's attack suppression at noticeably higher clean accuracy.
The claim is falsifiable. If a model trained with only the "lower" part, with the "raise" part fully suppressed, still learns the backdoor, the hypothesis is wrong.
A second, smaller observation is also new. Poisoning robustness only requires the add direction of DP. We need the probability that the trained model exhibits the backdoor, after the attacker adds poison set $P$, to be bounded by a multiple of that probability without $P$. We never need the reverse inequality. Ding et al.'s Asymmetric Differential Privacy formalizes neighbor relations that hold in only one direction under an application policy. Individualized Privacy Accounting via Subsampling with Applications in Combinatorial Optimization shows one-sided add-DP guarantees can be much cheaper than two-sided ones. Neither has been applied to training-time poisoning. We will report add-only accounting alongside the usual symmetric accounting, to measure how much of the budget is spent protecting a direction the attacker never uses.
Research Program
Write $J_i$ for the Jacobian of the model's class scores (logits) with respect to the parameters $\theta$ on example $i$. Write $p_{i,k}$ for its predicted probability of class $k$ and $y_i$ for its label. The per-example cross-entropy gradient splits exactly as
$$ g_i \;=\; \underbrace{-(1-p_{i,y_i})\,J_{i,y_i}^\top}_{u_i\ \text{(raise the label)}} \;+\; \underbrace{\textstyle\sum_{k\neq y_i} p_{i,k}\,J_{i,k}^\top}_{d_i\ \text{(lower the others)}} . $$
We will train with a two-channel update:
$$ \Delta\theta \propto \sum_i \mathrm{clip}(u_i, C_u) + \mathcal{N}(0,\sigma_u^2 C_u^2 I) \;+\; \sum_i \mathrm{clip}(d_i, C_l) + \mathcal{N}(0,\sigma_l^2 C_l^2 I). $$
Here $\mathrm{clip}(v, C)$ rescales $v$ to norm at most $C$. The terms $C_u, \sigma_u$ are the clipping bound and noise multiplier on the raise channel, and $C_l, \sigma_l$ are those on the lower channel. Each channel is a subsampled Gaussian mechanism, so its privacy loss $\epsilon_u$ or $\epsilon_l$ can be accounted separately, and the model as a whole satisfies their composition.
Setting $C_u = C_l$ and $\sigma_u = \sigma_l$ roughly recovers ordinary DP-SGD with a doubled per-example sensitivity. The regime of interest keeps $\epsilon_u$ strict and lets $\epsilon_l$ grow large or unbounded. Computing $u_i$ and $d_i$ requires two per-example backward passes, so the exact version costs about the same as DP-SGD, and we will say so plainly.
We will also test a cheap proxy that clips the raise part of the logit-space residual $p_i - e_{y_i}$ per example, with no per-example parameter gradients. It is essentially a loss reweighting and carries no formal guarantee. Comparing it with the exact version tells us whether the effect lives in logit space or requires parameter-space control.
The theory component is deliberately modest. We will derive add-only group bounds for the raise channel under Poisson subsampling. We will also check whether the add and remove directions of the subsampled Gaussian give materially different $\epsilon$ at the noise levels used. We will state openly that group-privacy certificates for realistic poison counts, around 1% of the data, are likely vacuous. The empirical question is the point.
Instance-Specific Asymmetric Sensitivity in Differential Privacy names the extension of directional sensitivity to high-dimensional gradients as open. Our decomposition does not solve that general problem. It supplies one structured direction, the class-score direction, that the loss itself defines.
Baselines
The naive reference point is undefended standard training on the same poisoned data. The file gives BadNets' 99% attack success rate (ASR) on MNIST with 99.5% clean accuracy (BadNets), and an undefended ASR of 97.12% at 1% poisoning in Robust anomaly detection and backdoor attack detection via differential privacy. The latter paper uses symmetric DP-SGD as a regularizer and anomaly scorer and drives ASR to "less than 1%".
The most direct competitor is symmetric DP-SGD itself, as evaluated in Does Differential Privacy Prevent Backdoor Attacks in Practice?. That paper compares DP-SGD, PATE and label-DP against BadNets with ResNet50 fine-tuning on CIFAR-10 and reports 88.67% clean accuracy for DP-SGD. It is the strongest competitor on our exact question because our method is a restriction of it. We will run it at matched total noise, with $\sigma$ and clipping tuned over the same grid and GPU-hour budget as our two-channel method. We will report the full Pareto curve of clean accuracy against ASR, not single points.
The strongest reported number in the file on this setting belongs to label-DP (ALIBI) from the same paper: 89.53% clean accuracy with ASR ≈ 0% on CIFAR-10 BadNets. Beating it means matching ASR ≈ 0 on BadNets at higher clean accuracy under the paper's setup. More importantly, it means keeping low ASR on clean-label attacks, where label-DP has no mechanism to act.
Among non-DP defenses, Effective Backdoor Defense by Exploiting Sensitivity of Poisoned Samples (FCT) filters samples whose features jump under rotations and shifts. It reports CIFAR-10 ASR of 1.21% at about 92% clean accuracy. Anti-Backdoor Coreset Selection via Cumulative Entropy (ABCS) trains on a high-entropy coreset of about 55% of the data and reports defense effectiveness rate (DER) above 95% on CIFAR-10. Both will be run through the BackdoorBench codebase (BackdoorBench) with default settings on identical poisoned datasets. We do not expect to beat FCT's clean accuracy on static triggers. We will count it as a win if the two-channel method is within about two points of FCT on clean accuracy while clearly beating FCT and ABCS on the attacks they concede, namely Adaptive Blend and transformation-invariant WaNet-style triggers, and across architectures.
Two ablations isolate the key ingredient. The first is clipping without noise in both channels; the second is symmetric clipping at the smaller of the two bounds. If either matches the two-channel method, the asymmetry is not what matters.
Why not just…?
Why not just use label-DP, which already gets ASR ≈ 0 at 89.53% accuracy? Label-DP randomizes labels, so it can only disrupt attacks that rely on flipped labels. In clean-label backdoors, the trigger is placed on correctly labeled target-class images, and label randomization leaves the trigger-to-target association intact. The raise channel still sees that association, because those images still push the target score up. The clean-label and Adaptive Blend arms of our experiments settle this directly. If ALIBI holds up there too, label-DP is sufficient and our method is unnecessary.
Why not just tune symmetric DP-SGD, or clip without noise? This is the real threat. Much of DP's anti-memorization effect may come from per-example clipping alone, and Robust anomaly detection and backdoor attack detection via differential privacy notes that the right noise scale is found by manual search. Our factorial design has four arms: clip versus no clip, noise versus no noise, symmetric versus asymmetric budgets, and matched tuning grids. It exists to answer exactly this question. If symmetric clipping at a well-chosen bound reaches the same Pareto front, we will report that as the finding and drop the asymmetric mechanism.
Why not just filter data with FCT or ABCS, which avoid DP's slowdown? They are faster and strong on static triggers. However, ABCS explicitly fails to neutralize Adaptive Blend and admits it lacks any explanation for why entropy separates poisons. FCT is defeated by triggers that survive the same transformations it uses. BackdoorBench reports that heuristic defenses tuned on ResNets fail on other architectures. A training-time mechanism does not rely on a detectable per-sample signature, which is where it should gain. If it does not, our adaptive-attack and cross-architecture arms will show it.
Why not just defend at inference with STRIP? STRIP superimposes clean images on each input and flags low prediction entropy. It reports 0% false acceptance and rejection in most CIFAR-10 and GTSRB cases at about 1.32x latency. But it leaves the backdoor in the model, pays that latency on every query, and its authors acknowledge difficulty with source-class-specific triggers. We will include source-specific attacks, where the question is whether the fixed model or the run-time filter degrades more.
Proposed Experiments
The main suite runs BackdoorBench's attacks, including BadNets, Blend, WaNet, a clean-label attack and Adaptive Blend. It covers CIFAR-10, GTSRB and Tiny-ImageNet with PreAct-ResNet18, plus MobileNet and DenseNet for the cross-architecture check. We add the ResNet50 fine-tuning BadNets setup of Does Differential Privacy Prevent Backdoor Attacks in Practice?, so that the label-DP and DP-SGD numbers are directly comparable. Metrics are clean accuracy, ASR, robust accuracy, DER, wall-clock time, and the accounted $\epsilon_u$ and $\epsilon_l$ under both symmetric and add-only accounting.
The first decisive check is the "lower-channel-only" run, with $C_u \to 0$. If backdoors are still implanted at high ASR through the lower channel alone, the upward-lift hypothesis is false.
The second check is the Pareto comparison against tuned symmetric DP-SGD and noise-free clipping at equal compute. The two-channel method must sit strictly above both curves on at least the clean-label and Adaptive Blend attacks.
The third check is an adaptive attacker who knows the mechanism. This attacker crafts poisons that maximize the lower-channel gradient toward the target, for example by optimizing triggers so that non-target scores fall rather than the target score rising. We report how much $\epsilon_l$ must tighten before ASR returns to the symmetric baseline. That threshold measures how much asymmetry is actually safe.
Finally, we report the gap between add-only and symmetric $\epsilon$ at the noise levels that suppress each attack, so the value of one-sided accounting is shown with a number rather than asserted.
References
- On the Lift, Related Privacy Measures, and Applications to Privacy-Utility Trade-Offs
- Asymmetric Local Information Privacy and the Watchdog Mechanism
- Asymmetric Differential Privacy
- Individualized Privacy Accounting via Subsampling with Applications in Combinatorial Optimization
- Instance-Specific Asymmetric Sensitivity in Differential Privacy
- BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain
- Robust anomaly detection and backdoor attack detection via differential privacy
- Does Differential Privacy Prevent Backdoor Attacks in Practice?
- Effective Backdoor Defense by Exploiting Sensitivity of Poisoned Samples
- Anti-Backdoor Coreset Selection via Cumulative Entropy
- BackdoorBench: A Comprehensive Benchmark of Backdoor Learning
- STRIP: a defence against trojan attacks on deep neural networks