Latent fingerprints are the partial, smudged prints recovered from crime scenes, often laid over wood grain, fabric, or dirt. The best deep matchers for them are now far ahead of classical tools. A dense-descriptor network reaches 79.07% rank-1 identification on the standard NIST SD27 benchmark, against 53.10% for a commercial SDK. These networks are also large and slow, and forensic galleries are huge. The obvious fix is to borrow post-training compression from large language models (LLMs): lower the numeric precision of the weights, or replace weight matrices with low-rank approximations, without retraining. Those methods decide what to keep by measuring gradients or activations on a small "calibration" set of inputs. On a latent print, most pixels are background texture, and that texture produces some of the strongest activations in the network. Compression calibrated on whole latent images may therefore spend its limited precision on the weights that respond to wood grain. The weights that read faint ridges, which are exactly what separates a hard match from a miss, would then be the ones rounded away.
This note proposes to measure that effect and correct it. We compute compression saliency and activation statistics only inside the regions of a print that hold recoverable ridge signal. Those regions are defined by the network's own foreground mask and by a local ridge-coherence quality measure. The claim is narrow and testable. At aggressive bit-widths, compression damage to a latent matcher falls mostly on the worst-quality prints. Restricting calibration to recoverable ridge regions should remove much of that damage at the same bit budget, and that masking, rather than simply calibrating on latent data, is what does the work.
Core Claim
LLM compression research has already shown that where you measure importance matters. Task-Circuit Quantization (TACQ) scores each weight by the expected loss change from quantizing it, sharpened by weight magnitude. It keeps the top few percent in 16-bit and quantizes the rest with GPTQ, recovering 96% of unquantized MMLU performance at 3.1 bits on Llama-3-8B-Instruct. ASVD shows that plain singular value decomposition fails when a few activation channels are huge. It rescales each weight matrix by per-channel activation magnitudes before truncating, and cuts parameters by 10–30%. GuidedQuant and AIR push the same logic further by guiding compression with end-task loss. These papers supply our method. All of them calibrate on text, where every token is signal.
In latent fingerprints, much of each input is structured noise. The biometric literature says this outright. LQM splits prints into Distinct, Retrievable-Corrupt, and Non-Retrievable blocks using gradient-covariance coherence, which measures how consistently local gradients point in one direction. Latent Fingerprint Matching via Dense Minutia Descriptor (DMD) scores matches only inside the intersection of predicted foreground masks, because background corrupts the score. Our contribution is to carry that masking discipline from the matching score into the compression objective.
Let $\ell(\theta)$ be the matching loss of network weights $\theta$ on a genuine latent–tenprint pair. Let $a_c(p)$ be the activation of channel $c$ at spatial location $p$. Let $m(p)\in[0,1]$ be a weight map that is 1 on foreground Retrievable-Corrupt blocks, 0.5 on Distinct blocks, and 0 elsewhere (these weights are a tunable choice). We replace the ASVD channel scale with a ridge-masked version:
$$ S_{cc} = \Big(\tfrac{\sum_p m(p)\,|a_c(p)|}{\sum_p m(p)}\Big)^{\alpha}, $$
where $\alpha$ is ASVD's scaling exponent. For TACQ-style saliency, we backpropagate a loss whose descriptor-similarity terms are restricted to the same masked region, so the gradient reflects only how each weight affects ridge evidence. Everything else in TACQ and ASVD is unchanged. This keeps the comparison clean: any gain comes from where calibration looks, not from a new quantizer.
Research Program
The first stage is diagnostic, and it can falsify the premise on its own. We take DMD, a dual-branch ResNet-34, as the primary target because it has published latent results. We compress it with round-to-nearest and GPTQ at 8, 4, 3, and 2 bits, and with ASVD at parameter ratios from 0.9 down to 0.6. We report rank-1 and true-accept rate (TAR) at 0.1% false-accept rate, stratified by SD27's examiner labels of Good, Bad, and Ugly. If the accuracy lost to compression is spread evenly across quality classes, there is no degradation-specific problem, and the project reduces to ordinary compression.
We also use the 5,449 examiner-verified minutia correspondences from Ground truth and evaluation for latent fingerprint matching. With them we check whether compression worsens the bias that paper documents, where systems reject genuine minutiae far more often than they accept false ones. This is the specific failure we expect: faint ridges go missing, not spurious ones appearing. Separately, we record where activation magnitude concentrates, in background versus foreground ridge blocks, to test whether the outlier channels ASVD protects are driven by substrate texture.
The second stage builds ridge-masked TACQ and ridge-masked ASVD as defined above. Masks come from DMD's own segmentation branch. LQM coherence labels are computed with a fast single-pass version that drops LQME's iterative genetic enhancement loop. Calibration uses genuine pairs from NIST SD302, and testing uses SD27, so the two sets never overlap.
The third stage tests transfer to a transformer. We apply the same procedure to IFViT, a Siamese vision transformer whose authors name computational overhead as future work. There we compress its query/key/value and cross-attention projections. IFViT has no reported latent evaluation, so we evaluate it on FVC2004 DB1 and cross-sensor MOLF. This stage asks whether masking helps, or at least does no harm, when inputs are cleaner.
Baselines
The uncompressed reference is DMD itself. DMD reports 79.07% rank-1 and 80.23% TAR at 0.1% FAR on NIST SD27, and 52.68% rank-1 on SD302. We rerun it under our pipeline so compressed variants are compared against our own measured full-precision numbers. The floor on the useful side is VeriFinger's reported 53.10% SD27 rank-1 from the same paper. A compressed DMD that falls below it has no practical value.
The naive reference is uniform round-to-nearest quantization at each bit-width, applied equally everywhere. The strongest general quantizer is GPTQ with the same calibration pairs. The strongest competitors on this exact question are unmasked TACQ and unmasked ASVD, calibrated on the same SD302 latent pairs with the same outlier budget $p\%$, the same rank search, and the same number of calibration examples and gradient passes. The only difference is that the mask is removed. This ablation isolates the key ingredient.
A second ablation replaces the LQM-weighted mask with DMD's binary foreground mask alone. It tells us whether ridge-coherence quality adds anything beyond segmentation. A third calibrates unmasked methods on clean tenprints, the likely default in practice.
There is no reported result on Lacuna for any compressed latent fingerprint matcher, and none for TACQ or ASVD on vision models. So we must produce all compressed baselines ourselves. We will count the proposal as beating unmasked TACQ/ASVD if, at 3 bits or at ASVD ratio 0.7, it recovers at least a further third of the rank-1 gap to full precision on SD27 Ugly prints, holds on SD302, and keeps overall rank-1 within noise of the unmasked methods on Good prints. Significance will be assessed with paired bootstrap over probes, since SD27 has only 258 latents.
Why not just…?
Why not just use standard int8 vision post-training quantization? For 8-bit weights this may well suffice, and our diagnostic stage will show it. If int8 DMD matches full precision on every quality class, we will say so and confine our claims to 4 bits and below. That is the regime where TACQ shows naive methods collapse on language models and where memory savings over gallery-scale deployment become worth having. This alternative is a real threat above 4 bits, and we treat it as a boundary to be measured, not argued away.
Why not just calibrate on latent images instead of clean ones? This is the most serious competitor, because domain-matched calibration alone might fix most of the problem. It is exactly our main ablation: unmasked TACQ and ASVD calibrated on the same SD302 latents. If they match the masked versions on Ugly prints, the masking claim is false and we would report domain-matched calibration as the practical recommendation.
Why not just triage with a global quality score and compress only the model used on good prints? LQM finds that global scores such as NFIQ miss localized core damage. Forensic practice also cannot discard bad latents, because the Ugly prints are where the evidence is hardest to recover. Routing systems from LLM serving, such as SymRAG and the teacher–student cache OCaTS, save cost only on easy or repeated queries. Unique latent prints give caching little to reuse, and a cheap path for hard prints is exactly what fails. We will still report a two-model triage (full DMD for Bad/Ugly, compressed for Good) as a reference point on the cost–accuracy curve.
Why not just fine-tune with quantization-aware training? Given enough labeled latent pairs it would likely do better. Labeled latent–tenprint pairs are scarce, and retraining gives up the no-retraining property that makes post-training methods deployable across vendors' matchers. We include a small quantization-aware fine-tuning run on SD302 as an upper reference, not as a competitor.
Proposed Experiments
We will plot rank-1 and TAR at 0.1% FAR against model memory and against latency on one GPU and one CPU, for DMD on SD27 (split into Good, Bad, and Ugly) and on SD302, across all methods and bit-widths. We will also measure memory for gallery descriptors, since DMD's dense per-location descriptors dominate storage at scale. For those we will test the same masked-activation low-rank projection applied to stored descriptors.
The decisive check is the interaction between method and quality class. Masking should help most on Ugly prints and least on Good prints, and the benefit should grow as bit-width falls. If the gain is uniform across quality classes, the mechanism is not what we claim, even if average numbers improve.
A second check uses the minutia-level ground truth. Masked compression should cut the rise in rejected genuine correspondences that unmasked compression causes, without raising false accepts. A third check applies the method to IFViT on FVC2004 DB1 and MOLF. Against IFViT's reported 1.87% and 2.33% EER, masking should roughly match unmasked compression on these cleaner prints, which would show it costs nothing where there is little background to mask.
References
- Task-Circuit Quantization: Leveraging Knowledge Localization and Interpretability for Compression
- ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models
- GuidedQuant: Large Language Model Quantization via Exploiting End Loss Guidance
- Activation- and Influence-Aware Ranks (AIR): Function-Preserving SVD Compression for LLMs
- LQM: Localized Quality Measure for Fingerprint Image Enhancement
- Latent Fingerprint Matching via Dense Minutia Descriptor
- Ground truth and evaluation for latent fingerprint matching
- IFViT: Interpretable Fixed-Length Representation for Fingerprint Matching via Vision Transformer
- SymRAG: Efficient Neuro-Symbolic Retrieval Through Adaptive Query Routing
- Cache me if you Can: an Online Cost-aware Teacher-Student framework to Reduce the Calls to Large Language Models