Organizations that hold sensitive data, such as hospital records or private code, often want two things from a pretrained model at once. They want to adapt it to their data without the model memorizing individual records, and they want the adapted model to be small and fast enough to deploy. The first need is served by differentially private training, which clips each example's gradient and adds calibrated noise so no single record can noticeably change the result. The second is served by structured pruning, which deletes whole channels, attention heads, or feed-forward neurons so the remaining weight matrices are smaller and still dense, and therefore genuinely faster on ordinary GPUs and CPUs. Today these two needs are met in sequence and at odds with each other. The pruning methods that know which structures matter for a task, such as LoRAPrune, read raw task gradients and so leak the private data. The private methods that fine-tune well, such as bias-only tuning, leave the full-size model in place.
We propose to fold the deletion decision into the private training run itself. A learnable scalar gate is attached to every prunable unit (each convolutional channel, attention head, and feed-forward neuron), and these gates are trained, along with biases only, using differentially private gradient descent and a sparsity penalty that pushes unneeded gates to exactly zero. Units whose gates end at zero are physically removed. Two features make this a good fit for privacy. The number of gates is tiny, on the order of the number of bias terms, so the injected noise is spread over very few dimensions. Deletion is a function of an already-private output, so the compact model inherits the privacy guarantee with no separate budget spent on choosing what to prune. In short, private gate training is at once the fine-tuning step and the pruning criterion.
Core Claim
Our claim is that per-unit gates trained privately on the sensitive task select structures to delete better than pruning on public data. The advantage should be largest when the private data's distribution differs from the public data. At the same time, the gates add almost no noise or memory cost beyond bias-only private tuning. The result should be a model that is meaningfully faster at inference, with accuracy close to private tuning of the unpruned model, under the same privacy budget.
Concretely, every prunable unit $c$ in layer $\ell$ gets its output $h_c$ replaced by $s_c h_c$, where $s_c$ is a scalar gate initialized to 1 and the pretrained weights stay frozen. We minimize
$$\min_{s,\,b}\; \frac{1}{N}\sum_{i=1}^{N} \mathcal{L}(x_i; W_0, s, b) \;+\; \lambda \sum_{c} \kappa_c\, |s_c|,$$
where $W_0$ are the frozen pretrained weights, $b$ are the bias terms, $\lambda$ sets overall sparsity, and $\kappa_c$ is the measured latency cost of unit $c$ from a hardware lookup table of the kind built by ZipLM. The data term is optimized with clipped, noised per-example gradients. The penalty term is applied as a soft-thresholding step after each noisy update. Because that step touches only parameters and never data, it costs no privacy. The per-example gradient of a gate is a single inner product, $\partial \mathcal{L}/\partial s_c = \sum_t \delta_{c,t} h_{c,t}$, where $\delta_{c,t}$ is the backpropagated error at position $t$. Storing it per example therefore costs memory proportional to batch size times number of units, not to the size of the weight matrices.
Relation to Prior Work
Three groups of papers bear on the idea, each in a different role.
The first group supplies the mechanism. Network Slimming puts an ℓ1 penalty on batch-norm scale factors and deletes channels whose scale goes to zero, reporting a 20× parameter reduction on VGGNet for CIFAR-10. That is our gate idea in a non-private, full-training setting, and we move it into private, frozen-backbone adaptation. DP-BiTFiT shows that privately tuning only biases is cheap and accurate: it reaches 88.3% on MNLI with RoBERTa-large at ε=8, against 87.0% for private full fine-tuning. Gates are the multiplicative counterpart of biases, so our method is DP-BiTFiT plus one scalar per unit. That scalar turns a tuning method into a pruning method.
The second group motivates the gap. SPARTA shows that grouping noisy gradients by row lets a coherent signal rise above DP noise. It spends one epoch of budget choosing which rows to tune, reaching 69.58% on CIFAR-100 with DeiT-Tiny at ε=2 versus 61.22% for private full fine-tuning. However, SPARTA freezes the unselected rows rather than deleting them, so its deployed model is full size. We aggregate at the level of whole units, which are coarser than rows and are exactly what structured pruning removes. Our selection also comes out of training itself rather than a separate budgeted phase. LoRAPrune does delete heads and channels, cutting LLaMA-7B inference time by 49.5%, but it uses unclipped adapter gradients. SPruFT tunes structured rows to save activation memory, also without privacy. Dual-Priv Pruning prunes visual tokens and noisy gradient blocks but never shrinks the weights.
The third group constrains the design. The DP sparse-training papers by Adamczewski and colleagues, Differential Privacy Meets Neural Network Pruning and Pre-Pruning and Gradient-Dropping, show that shrinking the set of parameters that receive noise helps accuracy. They use unstructured masks, which do not speed up dense matrix hardware, as argued in Learning Intrinsic Sparse Structures within LSTM. They also report that pre-pruning on public data degrades when that data is unlike the private data, which is the failure our method targets. Finally, MineGrad shows that adapter gradients can be inverted to reconstruct up to 85.9% of image patches without formal privacy. This rules out relying on the claimed natural filtering of low-rank updates reported by Das et al., and it is why every component here, including the pruning decision, sits under DP accounting.
Research Program
The method has three parts. First, we insert gates into the prunable units in a way that respects dependencies between layers. Following Network Slimming's lesson about residual networks, one gate is shared across every channel that must be pruned together to keep tensor shapes aligned. In transformers, one gate covers each attention head and each feed-forward neuron, the latter shared across the gate, up, and down projections as in 2SSP.
Second, we run a single private training phase on gates and biases. Proximal soft-thresholding produces exact zeros, and the final deletion threshold is set relative to the known noise scale. Because the noise standard deviation per step is public, we can delete a gate only if its averaged trajectory over the last epochs lies within a few noise standard deviations of zero. Doing this means a gate is not deleted merely because noise happened to push it toward zero at the end of training. The averaging is again post-processing.
Third, we optionally spend a portion of the same total budget on a recovery phase on the shrunk model, using private LoRA in the style of Yu et al. or continued bias tuning. Privacy loss across both phases is composed and reported as a single (ε, δ).
The analysis asks three questions. Where does the method's advantage come from: the private data signal, the low noise dimension, or simply the sparsity penalty? How does the accuracy-versus-latency curve shift as ε falls below 1, where Yu et al. report private adaptation degrading? Does deleting units change empirical memorization, measured with the loss-difference scores of Rare, Distinctive, Memorized? The last question is a secondary probe of the tension between low-capacity updates acting as a filter and acting as an attack surface. It does not replace the formal guarantee.
Baselines
The naive reference point is DP-BiTFiT on the unpruned model, as in Bu et al.. It gives the accuracy ceiling at zero speedup, with reported values of 88.3% on MNLI (RoBERTa-large, ε=8) and 92.3% on CIFAR-100 (ViT-large). We report our method as the accuracy lost relative to this point at each measured speedup. Success means staying within one point of it at a measured 1.5× inference speedup on both an A100 and a CPU.
The strongest practical competitor is pruning on public data followed by private adaptation. Concretely, we run 2SSP for transformers or Network Slimming for CNNs on public calibration data (ImageNet for vision models, C4 for language models), hitting the same latency target from the same lookup table. We then privately adapt the result with DP-LoRA or DP-BiTFiT. This competitor is strong because it spends nothing from the private budget on selection. It gets the same private epochs, the same ε, and the same hyperparameter search size as our method. To beat it, our method must achieve higher accuracy at matched measured latency, with the gap expected mainly on the shifted task. No reported result was found on Lacuna for any differentially private, physically pruned model, so this comparison is new in both directions.
The strongest existing private method is SPARTA. We run it as published and also in a variant that deletes, rather than freezes, the rows outside its mask at matched latency. The second variant tests whether a private mask originally chosen for tuning can serve as a deletion mask. SPARTA's reported 69.58% on CIFAR-100 (DeiT-Tiny, ε=2) anchors this comparison. We also implement a private version of LoRAPrune, which clips and noises adapter gradients and computes Taylor importance from them. It shows what happens when the natural existing criterion is simply privatized.
Three ablations isolate the key ingredient. The first trains gates privately without the sparsity penalty and prunes by final gate magnitude, which tests whether the penalty matters. The second trains gates on public data and only biases privately, which tests whether the private signal matters. The third removes the noise-aware threshold.
Why not just…?
Why not just prune on public data and then run private LoRA? This is the real threat. If public and private data match, as with ImageNet versus CIFAR-10, public pruning may do as well as our method, since it pays nothing from the privacy budget. The evidence against it as a general answer is the reported degradation of public pre-pruning under distribution shift in Differential Privacy Meets Neural Network Pruning. That evidence comes from unstructured pruning, so it is suggestive for our setting, not decisive. The experiment pairing near-domain and far-domain private tasks settles the question. If our method wins only on far-domain tasks, we will say so, and position it as the choice when no matching public data exists.
Why not just privatize LoRAPrune's adapter-gradient importance? That criterion places noise on all adapter parameters, which number orders of magnitude more than our gates. It must then compress them into one score per unit, and it needs separate budget for scoring. Our gates receive noise directly at the granularity of the decision. The private-LoRAPrune baseline measures whether this difference matters in practice. If it matches our method, the contribution shrinks to a cheaper implementation, and we will report it that way.
Why not just use a smaller pretrained model with DP-BiTFiT? When a smaller model from the same family exists at the target latency, as DeiT-Tiny does relative to DeiT-Small, this baseline costs nothing and could win. We include it at matched latency. Our method is useful when the smaller model is worse, when no model of the right size exists, or when a deployment target needs a latency point between the released sizes.
Why not just use unstructured private pruning? Earlier work shows unstructured sparsity rarely shortens wall-clock time on dense hardware (Compacting Deep Neural Networks for IoT). We report measured latency for the Pre-Pruning approach at matched sparsity to confirm this rather than assume it.
Proposed Experiments
The vision experiments use ImageNet-pretrained DeiT-Tiny, DeiT-Small, and ResNet models. The near-domain private tasks are CIFAR-10 and CIFAR-100. SVHN serves as a deliberately shifted private task, since digits are poorly represented in ImageNet. We test ε in {0.5, 1, 2, 8}, so results align with SPARTA's ε=2 numbers and the 94.84% at ε=1 reported for pre-pruned fine-tuning on CIFAR-10. The language experiments use RoBERTa-large on MNLI and the other GLUE tasks, plus GPT-2 on E2E and DART. The latter connects to the 43.8 BLEU that Yu et al. report for private GPT-2-XL. For every method we report task accuracy or BLEU, measured inference latency on an A100 and a CPU, peak training memory, and the composed (ε, δ).
The decisive check is the comparison against public pruning on SVHN and on the lowest-ε language tasks at matched measured latency. If private gates do not beat public pruning there, the central claim fails. A second check is that training memory for gates plus biases stays within a small factor of DP-BiTFiT, confirming that gates do not bring back the activation memory problem. The third check is the ablation that trains gates on public data. If it matches the full method, the private signal is not what selects the structures, and the idea reduces to a convenient implementation of public pruning.
References
- LoRAPrune: Structured Pruning Meets Low-Rank Parameter-Efficient Fine-Tuning
- ZipLM: Inference-Aware Structured Pruning of Language Models
- Learning Efficient Convolutional Networks through Network Slimming
- Differentially Private Bias-Term Fine-tuning of Foundation Models
- SPARTA: An Optimization Framework for Differentially Private Sparse Fine-Tuning
- An Efficient Row-Based Sparse Fine-Tuning
- Dual-Priv Pruning : Efficient Differential Private Fine-Tuning in Multimodal Large Language Models
- Differential Privacy Meets Neural Network Pruning
- Pre-Pruning and Gradient-Dropping Improve Differentially Private Image Classification
- Learning Intrinsic Sparse Structures within Long Short-Term Memory
- MineGrad: Gradient Inversion Attacks on LoRA Fine-Tuning
- Revisiting Privacy, Utility, and Efficiency Trade-offs when Fine-Tuning Large Language Models
- 2SSP: A Two-Stage Framework for Structured Pruning of LLMs
- Differentially Private Fine-tuning of Language Models
- Rare, Distinctive, Memorized: Auditing Memorization in Fine-Tuned Medical Foundation Models
- Compacting Deep Neural Networks for Internet of Things: Methods and Applications