GraFPrint: Multidomain Concepts & Applications

• GraFPrint is a versatile research label that defines graphene-based functional printing, GNN-powered audio identification, and biometric applications.
• In graphene electronics, GraFPrint enables flexible, transparent devices with quantified metrics like conductivity, thickness control, and percolation-managed transport.
• In audio and biometrics, it underpins a GNN-based audio fingerprinter and a Gram-feature extraction approach for fingerprint presentation attack detection, highlighting domain-specific innovations.

GraFPrint is a context-dependent research label rather than a single, universally fixed method. In the published record, it denotes a vision of graphene-based functional printing for flexible and transparent electronics, a graph-neural-network audio identification framework, a lightweight fingerprint presentation attack detection model based on Gram-feature extraction, and a fingerprint segmentation benchmark derived from FVC datasets with manually marked ground truth masks (Torrisi et al., 2011, Bhattacharjee et al., 2024, Park et al., 2018, Thai et al., 2015). By contrast, “GraphPrint” is a distinct name used for drug–target affinity prediction from 3D protein structure and should not be conflated with GraFPrint (Singh, 2024).

1. Terminological scope and naming ambiguity

The literature uses the label in several unrelated technical domains. In graphene electronics, GraFPrint denotes manufacturing flexible, transparent electronics entirely by printing graphene-based inks. In audio identification, GraFPrint expands to “Graph-based Audio Fingerprint” and names a specific GNN-based fingerprinter. In biometrics, GraFPrint appears both as a fingerprint liveness detector based on Gram modules and as the common community label for an FVC-based fingerprint segmentation benchmark, even though the segmentation paper itself does not explicitly use that label (Torrisi et al., 2011, Bhattacharjee et al., 2024, Park et al., 2018, Thai et al., 2015).

Domain	Meaning of GraFPrint	Primary source
Printed electronics	All-printed, flexible, transparent graphene electronics	(Torrisi et al., 2011)
Audio identification	GNN-based audio fingerprinting framework	(Bhattacharjee et al., 2024)
Fingerprint PAD	Gram-module CNN for live/spoof detection	(Park et al., 2018)
Fingerprint segmentation	FVC-based benchmark with ground-truth ROI masks	(Thai et al., 2015)

A persistent source of confusion is orthographic proximity to “GraphPrint.” The latter is explicitly a multimodal DTA framework using protein and drug graphs and traditional fingerprints, and its paper states that if “GraFPrint” is intended to refer to that work, the correct name is GraphPrint (Singh, 2024). This suggests that citation by label alone is insufficient; domain, task, and cited source are necessary to disambiguate the term.

2. GraFPrint as graphene-based functional printing

In printed electronics, GraFPrint denotes the use of printable graphene inks to realize flexible, transparent devices. Torrisi et al. established the practical foundation with a surfactant-free graphene ink produced by liquid phase exfoliation of graphite in N-methylpyrrolidone, printable with a modified Epson Stylus 1500 piezoelectric drop-on-demand system using a $\sim 50\,\mu\text{m}$ nozzle and $\sim 10\,\text{nL}$ drop volume (Torrisi et al., 2011). The work reports $Z \approx 1.7$ within the accepted stable jetting range $1 < Z < 14$, dried drop diameters of $\sim 90\,\mu\text{m}$ on HMDS, thermal removal of NMP at $170^\circ\text{C}$ for $5\,\text{min}$, and linearly increasing thickness with pass count, reaching a $90\,\text{nm}$ stripe after 30 passes. Transparent and conductive patterns with $T \approx 80\%$ and $R_s \approx 30\,\text{k}\Omega/\square$ were demonstrated, while printed graphene TFTs achieved $\sim 10\,\text{nL}$0 with $\sim 10\,\text{nL}$1 at room conditions; hybrid graphene/PQT-12 devices yielded $\sim 10\,\text{nL}$2 and $\sim 10\,\text{nL}$3 (Torrisi et al., 2011).

A central physical theme in this GraFPrint usage is percolation-controlled transport. For films thinner than $\sim 10\,\text{nL}$4, conductivity follows $\sim 10\,\text{nL}$5, with reported log–log exponents $\sim 10\,\text{nL}$6 on HMDS and pristine substrates and $\sim 10\,\text{nL}$7 on $\sim 10\,\text{nL}$8 plasma-treated glass (Torrisi et al., 2011). The same work ties improved conductivity to substrate engineering: HMDS reduces spreading, improves flake uniformity, promotes adhesion, and yields smoother networks with $\sim 10\,\text{nL}$9 rather than $Z \approx 1.7$0–$Z \approx 1.7$1.

Later work broadened the materials and manufacturing envelope. A water-based ink made from electrochemically exfoliated graphene was formulated in less than 5 h at a concentration of $Z \approx 1.7$2, remained stable for $Z \approx 1.7$3 month, contained more than $Z \approx 1.7$4 single- and few-layers with a log-normal thickness peak at $Z \approx 1.7$5 and mean lateral size of $Z \approx 1.7$6, and achieved $Z \approx 1.7$7 after a $Z \approx 1.7$8, 1 h anneal in $Z \approx 1.7$9 (Parvez et al., 2019). The same study used a Fujifilm Dimatix DMP-2800 with $1 < Z < 14$0 nozzles, reported $1 < Z < 14$1, and nevertheless observed stable jetting without satellite drops or clogging (Parvez et al., 2019). This indicates that empirical printability of graphene dispersions can extend beyond the classical $1 < Z < 14$2 heuristic when waveform tuning and substrate choice are favorable.

GraFPrint has also been pushed toward RF devices. Screen-printed and spray-coated exfoliated graphite dipole antennas for passive UHF RFID were fabricated on Kapton and paper, using a hybrid topology in which a small Al loop both matches the predominantly capacitive RFID chip and avoids direct FLG–chip bonding (Jaakkola et al., 2019). The hybrid tag reached a measured peak reading distance of $1 < Z < 14$3, close to the simulated $1 < Z < 14$4, with $1 < Z < 14$5 near 915 MHz and nearly zero mismatch loss (Jaakkola et al., 2019). A related, but process-distinct, extension is low-temperature dry transfer-printing of patterned graphene monolayers onto PEDOT:PSS and $1 < Z < 14$6 with a target-side thermal budget of $1 < Z < 14$7, enabled by a 200 nm Au support layer and a $1 < Z < 14$8 vol% ethanol–water bath with surface tension $1 < Z < 14$9 (Cha et al., 2015).

Across these graphene-printing uses, GraFPrint denotes a fabrication agenda rather than a single canonical architecture. The recurring constraints are rheological window, nozzle compatibility, percolation-limited conductivity, substrate wettability, and low-temperature post-processing.

3. GraFPrint as GNN-based audio identification

In audio identification, GraFPrint is a specific framework titled “GraFPrint: A GNN-Based Approach for Audio Identification” (Bhattacharjee et al., 2024). The system converts audio sampled at 16 kHz into a log-power Mel spectrogram of size $\sim 90\,\mu\text{m}$0, concatenates time and frequency index channels to form $\sim 90\,\mu\text{m}$1, applies strided 2D convolutions, projects each time–frequency point with a $\sim 90\,\mu\text{m}$2 convolution, constructs a dynamic undirected k-NN graph in latent space, and processes it with max-relative graph convolutions (Bhattacharjee et al., 2024). The core aggregation is defined as

$\sim 90\,\mu\text{m}$3

followed by linear update and nonlinear blocks, with average pooling over nodes and a fully connected projection to a 128-dimensional fingerprint vector $\sim 90\,\mu\text{m}$4 (Bhattacharjee et al., 2024).

Training is self-supervised and contrastive. GraFPrint uses NT-Xent loss with cosine similarity and temperature $\sim 90\,\mu\text{m}$5, together with augmentations for time offset, background noise mixing using MUSAN, and convolutional reverberation using Aachen RIRs (Bhattacharjee et al., 2024). Reported optimization details include Adam, cosine learning rate decay, batch size $\sim 90\,\mu\text{m}$6, 400 epochs, and training on two NVIDIA A100 GPUs. Retrieval is performed with FAISS IVF-PQ, using 256 centroids and codebase size $\sim 90\,\mu\text{m}$7, followed by offset compensation,

$\sim 90\,\mu\text{m}$8

and sequence-level scoring by inner product across overlapping query segments (Bhattacharjee et al., 2024).

The empirical results position GraFPrint as a strong large-scale fingerprinter. On the fma-medium reference set under noise only, 1 s queries achieved 63.9% at 0 dB and 98.8% at 20 dB; with noise plus reverb, 1 s queries achieved 52.3% at 0 dB and 89.4% at 20 dB, while 5 s queries reached 93.3%–97.7% (Bhattacharjee et al., 2024). On the scaled fma-large reference set with noise and reverb for 1 s queries, GraFPrint yielded 42.7%–83.8%, compared with 22.2%–40.8% for AST+IVFPQ (Bhattacharjee et al., 2024). The encoder has about 18M learnable parameters, smaller than the cited transformer baseline AST at about 45M (Bhattacharjee et al., 2024).

Subsequent work used GraFPrint as a neural baseline in a broader robustness study of music fingerprinting. In that evaluation, GraFPrint was characterized as extending NAFP with a GNN and being trained from scratch; under a common FAISS-based evaluation pipeline, it achieved an overall track-level top-1 hit rate of 67.82% (Singh et al., 7 Nov 2025). The reported condition-wise profile is uneven: 97 under pitch-shift, 95 under low-pass filtering, and 96 under echo, but 15 under band-pass filtering and 17 under Encodec compression (Singh et al., 7 Nov 2025). On segment-level Pexeso evaluation, GraFPrint obtained Track F1 78.0, BBox F1 49.4, and Length F1 67.6 on pexafb_hard_small, rising to 81.3, 61.9, and 70.4 on pexafb_hard_medium (Singh et al., 7 Nov 2025). This later comparison situates GraFPrint as a strong from-scratch baseline whose main strengths lie in graph-based structural encoding and whose main vulnerabilities, at least in that study, lie in certain spectral distortions and neural-codec compression.

4. GraFPrint as fingerprint presentation attack detection

A separate biometric use of the name denotes an end-to-end fingerprint liveness detector that embeds Gram-feature extraction inside a compact CNN (Park et al., 2018). This GraFPrint operates directly on raw grayscale fingerprint images of arbitrary size, without segmentation or resizing, and uses about 308,554 parameters, approximately 1.2 MB (Park et al., 2018). The backbone is built from SqueezeNet-style fire modules, while three Gram modules are inserted at different depths to extract multi-scale texture statistics. Each Gram module projects channels with a $\sim 90\,\mu\text{m}$9 convolution to $170^\circ\text{C}$0, applies $170^\circ\text{C}$1, reshapes to $170^\circ\text{C}$2, and computes a Gram matrix

$170^\circ\text{C}$3

which is then treated as a $170^\circ\text{C}$4 tensor (Park et al., 2018). Concatenating the three outputs along the channel axis yields a $170^\circ\text{C}$5 texture representation that feeds a compact classification head.

The design premise is that texture is the most appropriate characteristic in fake fingerprint detection. Gram matrices capture second-order correlations between feature channels aggregated over spatial locations, making the representation holistic, size-agnostic, and explicitly texture-oriented (Park et al., 2018). Training uses Adamax with initial learning rate 0.0005, batch size 8, 80 epochs, batch normalization, LeakyReLU with $170^\circ\text{C}$6, and a 10% validation split; the paper reports no preprocessing beyond raw grayscale input and experiments with random horizontal and vertical flips (Park et al., 2018).

Performance was evaluated on LivDet 2011, 2013, and 2015 using Average Classification Error,

$170^\circ\text{C}$7

The reported average ACE across those datasets is 2.61%, compared with 2.34% for the cited state of the art, while retaining the much smaller parameter footprint (Park et al., 2018). Representative ACE values include 0.55% on LivDet 2011 Digital Persona, 0.85% on LivDet 2013 Biometrika, and 0.27% on LivDet 2015 Crossmatch (Park et al., 2018). In this usage, GraFPrint does not refer to graph neural networks or graphene printing; it denotes Gram-based fingerprint liveness detection.

5. GraFPrint as a fingerprint segmentation benchmark

In fingerprint segmentation, GraFPrint refers to a benchmark comprising 10,560 fingerprint images from FVC2000, FVC2002, and FVC2004 with manually marked ground-truth foreground/background segmentation masks (Thai et al., 2015). The benchmark uses 12 databases, each with 880 images; for every database, 80 images are used for training and 800 for testing, giving 960 training images and 9,600 test images overall (Thai et al., 2015). The paper employing it states that, in the fingerprint segmentation literature, this dataset is widely referred to as the GraFPrint benchmark, although the paper itself does not explicitly use the label (Thai et al., 2015).

The evaluation metric is average total pixel error per image,

$170^\circ\text{C}$8

where $170^\circ\text{C}$9 counts foreground pixels misclassified as background and $5\,\text{min}$0 counts background pixels misclassified as foreground (Thai et al., 2015). On this benchmark, the proposed G3PD method formulates segmentation as a global three-part decomposition into cartoon $5\,\text{min}$1, texture $5\,\text{min}$2, and residual/noise $5\,\text{min}$3, with total variation on $5\,\text{min}$4, curvelet-domain and spatial sparsity on $5\,\text{min}$5, and a bounded-noise constraint in curvelet space (Thai et al., 2015). Morphological post-processing of the texture component then produces the ROI mask.

The benchmark’s significance lies in standardized cross-sensor comparison. On the reported protocol, G3PD achieved the lowest average error, 3.06%, compared with 3.30% for FDB, 5.78% for HCR, 6.51% for MVC, 7.57% for STFT, and 8.33% for GFB (Thai et al., 2015). G3PD won on 10 of the 12 databases, while FDB was best on FVC2000 DB1 and DB2 (Thai et al., 2015). In this usage, GraFPrint denotes not an algorithmic architecture but an evaluation resource for foreground/background segmentation.

6. Cross-domain patterns, misconceptions, and citation practice

The distinct GraFPrint usages share very little at the level of task or mechanism. In graphene electronics, the label refers to printable graphene materials, jetting physics, percolation, and device fabrication (Torrisi et al., 2011). In audio identification, it names a graph-structured neural fingerprinter over time–frequency features (Bhattacharjee et al., 2024). In fingerprint biometrics, it has been used for Gram-based presentation attack detection and, separately, as a shorthand for an FVC-derived segmentation benchmark (Park et al., 2018, Thai et al., 2015). A plausible implication is that GraFPrint functions more as a locally meaningful project name than as a globally unique technical term.

Several adjacent names reinforce the need for disambiguation. “GraphPrint” is the correct name for the protein-structure DTA framework and not a variant spelling of GraFPrint (Singh, 2024). “DrawnApart” is a remote GPU fingerprinting technique that explicitly states it does not mention or evaluate any system named “GraFPrint” (Laor et al., 2022). “DiffusionPrint” discusses generative fingerprints for diffusion-based inpainting localization and notes that its comparison to “GraFPrint” is conceptual rather than tied to an explicitly cited paper-specific method (Giakoumoglou et al., 14 Apr 2026). In audio fingerprinting, later work treats GraFPrint as a concrete baseline and compares it against foundation-model-based systems such as MuQ and MERT (Singh et al., 7 Nov 2025).

For citation practice, the least ambiguous convention is to pair the name with its domain and source: GraFPrint for graphene printing (Torrisi et al., 2011), GraFPrint for audio identification (Bhattacharjee et al., 2024), GraFPrint for fingerprint liveness detection (Park et al., 2018), or the GraFPrint fingerprint segmentation benchmark (Thai et al., 2015). Without that domain qualifier, the term is underspecified in contemporary arXiv-indexed usage.