Pith. sign in

REVIEW 6 major objections 5 minor 2 cited by

One Video to Steal Them All: 3D-Printing IP Theft through Optical Side-Channels

T0 review · 6 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A video of a 3D print is enough to recover the printable G-code, including layer boundaries, extrusion rate, and feed rate; the paper demonstrates a 90.87% average similarity and functional counterfeit parts.

desk verdict First plausible end-to-end optical side-channel to printable G-code, but the key-demo confound (manually measured bitting) undercuts the strongest functional claim. read the letter →

arxiv 2506.21897 v1 pith:FPRO7X6W submitted 2025-06-27 cs.CR

Twisha Chattopadhyay , Fabricio Ceschin , Marco E. Garza , Dymytriy Zyunkin , Animesh Chhotaray , Aaron P. Stebner , Saman Zonouz , Raheem Beyah This is my paper
classification cs.CR
keywords 3DprintingG-codereverseengineeringopticalside-channelintellectualpropertytheftnozzletrajectoryequivalencecheckercounterfeitparts
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the video feed of a 3D printing process is itself enough to steal the design's print instructions, without touching the printer, the network, or the design file. The authors' model watches the nozzle through the camera, recovers its path as G-code commands, and then fills in the parameters that make those commands actually printable: which movements extrude filament, how much filament each move pushes, the feed speed, and the 0.3 mm layer boundaries. The result, they claim, is the first optical side-channel attack that produces a complete, printable G-code rather than a rough trajectory. They support the claim with an average 90.87% similarity between recovered and original instructions on their dataset, plus two physical counterfeits—a padlock key that opens its lock and a 16-tooth gear that fits into a gear train—printed from recovered G-code. A reader should care because remote monitoring cameras are already common in printing facilities, so the attack surface is the equipment that operators believe is purely for quality control.

What carries the argument

The carrying mechanism is a two-stage pipeline with a companion metric. First, a video model (ResNet-50 frame encoder plus LSTM temporal encoder) converts each 30-frame clip into a predicted G0/G1 command and destination coordinates for one nozzle movement. Second, a deterministic post-processor turns those predictions into printable G-code: extrusion amount from the geometry formula $E=(4 h s l d_n)d_f^2/\pi$, feed rate from the command type, and layer boundaries from a PELT change-point detector applied to predicted Z values (each detected change increments Z by 0.3 mm). The argument that the recovered G-code is equivalent to the original rests on the G-code equivalence checker: it reduces both programs layer by layer to skeletal (X,Y) trajectories, aligns them with convex-hull oriented bounding polygons and a maximum-overlap rotation search, and scores the residual difference with subsequence-aligned dynamic time warping.

What would settle it

Print an object with a non-concentric infill (e.g., gyroid) or with a 0.2 mm layer height while recording with the same camera setup, run the published model, and compute the curve-checker similarity to the original G-code. If the similarity falls well below the reported 90.87% or the printed copy fails a functional test, the claim that a fixed-camera video alone is sufficient would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that a fixed, unmodified camera—the kind already used for remote monitoring—is a sufficient side channel to reverse-engineer a 3D-printable G-code IP. The paper's model takes 30-frame chunks of the print video, encodes each frame with a ResNet-50 feature extractor, and passes the sequence through an LSTM whose output feeds two heads: one classifies the movement as a travel command (G0) or an extrusion command (G1), and the other regresses the destination coordinates (X, Y, Z). The remaining fields are recovered deterministically rather than learned: the extrusion amount for a G1 move follows $E=(4 h s l d_n)d_f^2/\pi$ from the layer height $h$, flow $s$, line length $l$, nozzle diameter $d_n$, and filament diameter $d_f$; the feed rate follows from the predicted command type; and the Z-axis is discretized into 0.3 mm layers by applying a PELT change-point detector to the predicted Z values. To judge whether the recovered G-code really matches the original, the paper builds a rotation- and translation-invariant equivalence checker: each layer is reduced to its skeletal (X,Y) trajectory, enclosed in a convex-hull oriented bounding polygon, aligned by centroid shift and a maximum-overlap rotation search, and scored with subsequence-aligned dynamic time warping. On the paper's dataset the recovered G-codes score 90.87% average similarity and contain 30.20% fewer instructions than the prior ResNet-50 based recovery, and the two printed counterfeits pass functional tests (the key opens the padlock; the gear meshes in the gear train).

Load-bearing premise

The chain breaks if the victim's printer and slicer settings differ from the ones the attacker assumes—layer height 0.3 mm, nozzle diameter 0.4 mm, filament diameter 1.75 mm, flow 100%, concentric infill—because the extrusion amounts and Z-layer thicknesses are computed from those numbers rather than measured.

Editorial extensions

If this is right

  • An attacker who can watch a print through the facility's own monitoring camera can produce a working counterfeit part without installing any device or compromising the printer.
  • Recovered G-code is not just visually similar; in the demonstrated cases it is functionally interchangeable, since the key opens the lock and the gear meshes with the gear train.
  • Defenses that degrade the optical channel, such as adding noise or changing lighting, are not viable in a commercial setting because they would also cripple the legitimate remote monitoring the camera exists for.
  • The recovered parameters (extrusion and feed rates) transfer across printers, so the attacker can print on a different machine than the one filmed.
  • The rotation- and translation-invariant equivalence checker provides a common yardstick both for attackers testing theft success and for defenders comparing suspected counterfeit G-code against proprietary files.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The evaluation is limited to 16 simple open-source objects printed on one printer with one slicer configuration; the authors do not show that the same accuracy holds for large, densely infilled industrial parts, and their own discussion notes complex-object evaluation as future work.
  • Because extrusion and layer-height reconstruction are computed from assumed slicer settings, a victim who randomizes layer height, flow, nozzle diameter, or infill pattern per job would likely break the parameter-derivation step even though the outer toolpath would still be visible.
  • The same 'watch the toolpath, reconstruct the program' logic may apply to other visible-toolpath digital fabrication processes, such as CNC machining or laser cutting, where a fixed camera sees the head move.
  • The curve checker, being invariant to rotation and translation, could be reused as a general-purpose G-code similarity tool outside this attack—for example, to detect whether a G-code file was derived from another, regardless of where the workpiece was placed on the bed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. This paper proposes a side-channel attack that recovers a printable G-code from a video of a fused-filament 3D printing process. The pipeline uses a ResNet-50+LSTM model to predict, from 30-frame video chunks, the movement command (G0/G1) and the destination (X,Y,Z) coordinates; a sliding window produces a full trajectory. Z-values are discretized into layers using PELT change-point detection, extrusion lengths are computed from a deterministic equation, and feed rates are assigned from a command-type lookup. The authors also introduce an equivalence checker that aligns per-layer trajectories via centroid translation, a convex-hull 'oriented bounding polygon,' and a 1-degree rotation search, then scores similarity with subsequence-aligned DTW. Evaluation on a 16-object dataset reports 90.87% average similarity and functional counterfeits of a padlock key and a 16-tooth gear.

Significance. If the video-only recovery claim were fully supported, this would be a valuable contribution to the side-channel literature, with implications for remote-monitoring security in additive manufacturing. The paper has notable strengths: the dataset is large and self-collected with timestamp synchronization; the gear counterfeiting test is an external functional benchmark that does not depend on the authors' checker; the equivalence checker is evaluated on rotated/translated variants and is markedly more stable than nMSE; and the authors are explicit about several limitations. However, the strongest functional demonstration (the key) is confounded by manual entry of the bitting, and the headline accuracy metric is not yet shown to track functional equivalence. These issues are fixable in a revision but are load-bearing for the paper's central claim.

major comments (6)
  1. [§5.2 (Counterfeiting via reverse-engineering), Abstract] The padlock-key demonstration does not establish that the bitting was recovered from video: the text states 'We manually recorded the bitting of the key that opens the padlock and used keygen to generate a 3D model with an equivalent bitting.' The bitting is the key's functional geometry, and it was provided to the pipeline before any video was processed. The recovered G-code is therefore judged against a G-code generated from the same manual measurement. To support the claim of video-only IP theft, the key experiment must be repeated with the bitting inferred solely from the video, or the claim should be scoped to non-bitting geometry.
  2. [§5.2 and §4.2 (Equivalence checker)] The checker similarity does not by itself measure functional equivalence. The paper reports that KE4 achieved 91.8% similarity but had 'incorrect bitting,' while the reused model output 91.8% similarity and was functionally correct after retraining on extra keys. This decoupling means the headline 90.87% average similarity cannot be read as evidence that functional geometry (e.g., key bitting or gear teeth) was stolen. The authors should either validate the checker against a panel of functionally tested objects or present the accuracy number as a trajectory-similarity score, not as an IP-theft success rate.
  3. [§1, §4.1, §5] The threat model states that the adversary has 'no knowledge of the original G-code' and does not know the parameters used to create it, but the pipeline requires those parameters: the extrusion equation uses fixed layer height 0.3 mm, flow 100%, nozzle diameter 0.4 mm, and filament diameter 1.75 mm; Z discretization increments by exactly 0.3 mm; and the evaluation fixes the infill pattern to concentric. Section 5 states these choices for the authors' printer and slicer, but a real target using different values would yield incorrect extrusion amounts and layer boundaries. The paper should either revise the threat model to include known printer/slicer configuration or add a sensitivity analysis over these parameters.
  4. [§4.1 (Obtaining the Extrusion and Feed Rate)] Feed rate is not predicted from the video; it is assigned by a two-entry lookup: 'most G0 commands have a feed rate of 7740, while G1 commands have a feed rate of 3600. Thus, based on the label given by the command prediction model, we can estimate the value of the feed rate F.' This supports the claim of identifying feed rate only for this specific dataset. The paper should either demonstrate feed-rate recovery on targets with other speed profiles or drop the feed-rate component from the contribution list.
  5. [§5.1 and §5.2 (Robustness to camera angles)] The robustness claim is internally inconsistent. Section 5.1 says G-code-to-video dataset #2 was recorded at a 60-degree angle for objects PYR, HPY, and SOC; Section 5.2 says the model was tested on HPT and HOC in dataset #2. If HPT and HOC were not in dataset #2, the experiment as described cannot be performed; if they were, the composition of dataset #2 is misreported. In addition, training on both camera angles before testing on a new angle does not test generalization to an unseen viewing angle. The authors should clarify the dataset composition and, ideally, evaluate on an angle not used in training.
  6. [§5.3 (Comparison with prior work)] The comparison with Liang et al. is not a faithful reproduction: the paper extracts the ResNet-50 from its own model as a 'starting checkpoint' and trains it on the last frame of each video. Since Liang et al.'s method was trained as an image-to-point regression on their own data, the reported loss curves and the '30.20' instruction-overhead factor do not necessarily characterize the original method. The comparison should be based on a published or clearly specified reimplementation, and the '30.20 percent' (abstract) versus '30.20 times' (Figure 16 caption) discrepancy should be resolved.
minor comments (5)
  1. [Abstract vs. §5.3 / Figure 16] The abstract reports '30.20 percent fewer instructions' while Section 5.3 and Figure 16 state '30.20 times fewer instructions'; these differ by two orders of magnitude and should be reconciled.
  2. [§4.1 (Normalizing the Z-Axis)] The PELT penalty cost and minimum segment length are described as selected by the attacker 'for a given 3D object,' but the selection procedure is not described; since these are free parameters in the reported accuracy, the paper should state how they are chosen.
  3. [§4.2 (Perfect Alignment)] The rotation alignment uses an exhaustive 1-degree search over 360 degrees; the paper should report the computational cost and whether a coarser search changes the final similarity scores.
  4. [Figure 15 and §5.3] Figure 15 compares loss curves of the proposed LSTM model and a ResNet-50 reproduction, but the reproduction is not a published implementation of Liang et al.; the figure should be labeled as a re-implementation on the authors' dataset.
  5. [§5.1 (Dataset #1)] The claim that this is 'the largest existing G-code-to-video dataset in literature' is plausible but should be supported by citing the dataset sizes of the closest prior works, e.g., video counts in Liang et al.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial circularity: feed-rate 'identification' is a dataset lookup and the functional-key demo is pre-seeded with manual bitting; core 90.87% trajectory recovery remains independent.

  1. fitted input called prediction [Section 4.1, 'Obtaining the Extrusion and Feed Rate']
    "In our experimental setup and dataset, most G0 commands have a feed rate of 7740, while G1 commands have a feed rate of 3600. Thus, based on the label given by the command prediction model, we can estimate the value of the feed rate F."

    The feed rate is not inferred from the video or from the stated extrusion equation; it is a two-value lookup fitted to the training dataset's command labels. The Abstract's claim that the model 'identifies the correct parameters such as feed rate' therefore reduces, for feed rate, to memorized dataset statistics. A target sliced with different feed rates would be mispredicted by construction, and the paper reports no per-instruction validation of F.

  2. other [Section 5.2, 'Counterfeiting via reverse-engineering']
    "We manually recorded the bitting of the key that opens the padlock and used keygen to generate a 3D model with an equivalent bitting."

    The key's functional geometry (bitting) is measured and supplied before any video is processed; the video is then of a print of that manually constructed model. The Abstract's headline 'a fully functional counterfeit object generated by reverse engineering 3D print instructions from video' is therefore not established for the key: the padlock-opening success is predetermined by the manually recorded bitting, not independently recovered from the video. The gear case is cleaner, but the key is the flagship functional demonstration.

full rationale

The core trajectory reverse-engineering is not circular: the ResNet+LSTM regresses coordinates from video, is evaluated on held-out objects (TET, DOD, G16, KE4, etc.), and the gear substitute is produced from the recovered G-code, an external functional check. The extrusion rate follows a stated deterministic formula from assumed printer parameters, and the equivalence checker is validated on rotated/translated variants (99.76%/99.71%/99.54%). However, three steps undercut the strongest claims: (1) feed rate is assigned from dataset frequencies (7740 vs 3600) rather than derived from video; (2) PELT layer segmentation invites per-object parameter selection, so layer counts are partly tuned; (3) the padlock-key demo is seeded with manual bitting, making the 'fully functional counterfeit from video' claim partially constructed. The comparison against Liang et al. [45] is also a self-citation (shared authors Zonouz and Beyah) reproduced from the current authors' own checkpoint, though this baseline comparison is not load-bearing for the main accuracy number. These are partial reductions of sub-claims, not a collapse of the whole derivation, hence a 4 rather than a 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central derivation rests on known printer parameters and infill choices rather than fitted constants. The only free knobs are the PELT layer-detection parameters, the feed-rate lookup values, and the sliding-window hyperparameters. No new physical entities are introduced; the OBP alignment and G-code manipulator are algorithms, not postulates.

free parameters (4)
  • PELT penalty cost and minimum segment length
    Z-axis layer segmentation uses PELT with parameters selected by the attacker per object, Section 4.1. This knob controls layer count and thus the recovered object height.
  • Feed rate lookup table per command type = G0: 7740, G1: 3600
    Section 4.1: feed rate is not predicted from the video; it is assigned from the predicted G0/G1 label using values observed in the authors' own dataset. Other printers and slicers use different speeds.
  • Sliding window size and stride = window=60 frames, sample=30 frames, stride=30 frames
    Section 4.1: these hyperparameters determine how many G-code instructions are produced and the temporal context; chosen by hand without ablation, and they drive the 30x fewer instructions comparison.
  • Rotation search step in equivalence checker = 1 degree
    Section 4.2: alignment searches 1 to 360 degrees at 1 degree increments; a coarser or finer step could affect similarity scores for trajectories whose best alignment falls between integer degrees.
assumptions (5)
  • domain assumption The 3D printer is a filament-based single extruder FDM printer running Marlin with known slicer settings (layer height 0.3 mm, nozzle diameter 0.4 mm, filament diameter 1.75 mm, flow 100%).
    Sections 3 and 5: the deterministic extrusion equation and Z-discretization require these values; an attacker facing a different printer or slicer config would need to know or infer them.
  • domain assumption The target objects are printed with concentric infill so that the infill pattern is identical regardless of build-plate position.
    Section 5 states this choice explicitly. The paper's own invariance evaluation (Section 5.4) shows non-concentric infill changes under rotation and translation, so the attack and checker are validated only for concentric infill.
  • domain assumption Camera view is fixed and shows the full printer without occlusion, and the video is synchronized to G-code execution via an inserted M300 tone.
    Threat model Section 1 and data pipeline Section 5.1: if the nozzle is occluded or the view changes, the sliding-window tracking fails.
  • domain assumption G0/G1 linear moves fully describe the printed object; arcs and non-linear moves are not considered.
    Section 3 restricts to linear movements supported by all printers; real slicers may emit G2/G3 arcs or other commands.
  • standard math PELT change-point detection is a valid and sufficient method for layer segmentation given the attacker can tune it per object.
    Section 4.1 uses PELT to identify Z-axis shifts; the method itself is standard, but its parameter selection is object-specific, making layer recovery partly manual.

how reviews work

0 comments
Cite this review

Pith. "Pith review of One Video to Steal Them All: 3D-Printing IP Theft through Optical Side-Channels." pith.science (2026). https://pith.science/paper/FPRO7X6W

@misc{pith2026250621897,
  author       = {Pith},
  title        = {Pith review of: One Video to Steal Them All: 3D-Printing IP Theft through Optical Side-Channels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FPRO7X6W}},
  note         = {Machine review of arXiv:2506.21897}
}
read the original abstract

The 3D printing industry is rapidly growing and increasingly adopted across various sectors including manufacturing, healthcare, and defense. However, the operational setup often involves hazardous environments, necessitating remote monitoring through cameras and other sensors, which opens the door to cyber-based attacks. In this paper, we show that an adversary with access to video recordings of the 3D printing process can reverse engineer the underlying 3D print instructions. Our model tracks the printer nozzle movements during the printing process and maps the corresponding trajectory into G-code instructions. Further, it identifies the correct parameters such as feed rate and extrusion rate, enabling successful intellectual property theft. To validate this, we design an equivalence checker that quantitatively compares two sets of 3D print instructions, evaluating their similarity in producing objects alike in shape, external appearance, and internal structure. Unlike simple distance-based metrics such as normalized mean square error, our equivalence checker is both rotationally and translationally invariant, accounting for shifts in the base position of the reverse engineered instructions caused by different camera positions. Our model achieves an average accuracy of 90.87 percent and generates 30.20 percent fewer instructions compared to existing methods, which often produce faulty or inaccurate prints. Finally, we demonstrate a fully functional counterfeit object generated by reverse engineering 3D print instructions from video.

Figures

Figures reproduced from arXiv: 2506.21897 by the authors.

Figure 1
Figure 1. Threat model. A 3D designer creates a 3D object and slices it to generate a G-code that can be printed by the 3D printer [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Infill Patterns: From left to right, the infill patterns [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Examples of G-code Instructions The primary action commands describe the type of movement the mechanical components of the 3D Printer is supposed to make from its current position to the destination position that is described using positional parameters. The primary action commands can be used for various tasks. The Gx commands are used to describe the movement type (e.g., straight-line, curve) of the nozzle aka pri… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: G-code Reverse Engineering Model. The model uses [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Model inference using a sliding-window strategy. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visual representation of two unique trajectories [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Layer 79 of S4 (given in Table 2) [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: The Data Collection Pipeline. Our strategy consists of sending the start G-code, followed by a tone, which is used to [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: All four images display Layer 840 of the Olympic [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Similarity between reverse-engineered G-code and [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Top: Reverse-engineered key trajectory overlapped over the original key’s trajectory in layer 3. Bottom: Function￾ality test of the counterfeit 3D-printed key. The counterfeit key is 91.8% similar to the original one and is able to open the padlock [PITH_FULL_IMAGE:f…
Figure 13
Figure 13. Figure 13: Left: Reverse-engineered gear trajectory over￾lapped over the original gear’s trajectory in layer 3. Right: Functionality test of a counterfeit 3D-printed 16-tooth gear. The counterfeit (middle) gear is 84.4% similar to the original one and is able to fit into the thr…
Figure 14
Figure 14. Figure 14: Camera Angles for Creating G-Code to-Video Datasets. Our main experiments used the camera angle shown in (a) to record the 3D printing process. To test the robustness of the model to camera angles, we changed the camera angle to 60◦ relative to (a) as shown in (b). no…
Figure 17
Figure 17. Figure 17: Z-Axis Normalization Using PELT Change Point [PITH_FULL_IMAGE:figures/full_fig_p013_17.png]
Figure 20
Figure 20. Figure 20: Translational and Rotational invariance of our [PITH_FULL_IMAGE:figures/full_fig_p014_20.png]
Figure 19
Figure 19. Figure 19: Translation invariance of our G-code Equivalence [PITH_FULL_IMAGE:figures/full_fig_p014_19.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Firewall3D: A Hardware Firewall for Defending 3D Printers Against Firmware Attacks

    cs.CR 2026-07 conditional novelty 6.5 of 10

    A dedicated hardware monitor detects firmware attacks on 3D printers by comparing real-time physical signals (stepper currents, temperatures, fans, endstops) against intended G-code execution.

  2. QuietPrint: Protecting 3D Printers Against Acoustic Side-Channel Attacks

    cs.CR 2026-02 conditional novelty 6.0 of 10

    Rewriting G-code to add decoy head movements can make an acoustic side-channel attacker reconstruct a different shape than the one being printed.

Reference graph

Works this paper leans on

72 extracted references · 57 canonical work pages · cited by 2 Pith papers

  1. [1]

    Marlin Firmware. 2024. What is Marlin? https://marlinfw.org/docs/basics/ introduction.html#main-features

    work page 2024

  2. [2]

    3D Universe. 2024. The Military Turns to 3D Printing. https://shop3duniverse. com/blogs/digital-fabrication-in-the-news/the-military-turns-to-3d-printing Accessed: 2024-05-30

    work page 2024

  3. [3]

    3DPrinterOS. 2024. 3D Printer Monitoring. https://www.3dprinteros.com/3d- printer-monitoring Accessed: 2024-11-18

    work page 2024

  4. [4]

    3DQue. 2018. 3DQue: Automated 3D Printing Solutions. https://www.3dque.com/. Accessed: 2024-11-18

    work page 2018

  5. [5]

    Mohammad Abdullah Al Faruque, Sujit Rokka Chhetri, Arquimedes Canedo, and Jiang Wan. 2016. Acoustic Side-Channel Attacks on Additive Manufacturing Systems. In 2016 ACM/IEEE 7th International Conference on Cyber-Physical Systems (ICCPS). 1–10. doi:10.1109/ICCPS.2016.7479068

    work page arXiv 2016

  6. [6]

    Hamza Alkofahi, Heba Alawneh, and Anthony Skjellum. 2024. MitM attacks on intellectual property and integrity of additive manufacturing systems: A security analysis. Computers & Security 140 (2024), 103810

    work page 2024

  7. [7]

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. 2021. ViViT: A Video Vision Transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . 6836–6846

    work page 2021

  8. [8]

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. 2021. ViViT: A Video Vision Transformer. arXiv:2103.15691 [cs.CV]

    arXiv 2021

Show all 72 references
  1. [9]

    Astroa. 2024. Guam to Host Innovative 3D Printing Center for Submarine Parts, Boosting Economy and Security. https://astroa.org/bnn-guam-to-host-innovative- 3d-printing-center-for-submarine-parts-boosting-economy-and-security/. Astroa News (2024)

    2024

  2. [10]

    AstroPrint. 2016. Live Monitoring for 3D Printing. https://www.astroprint.com/ live-monitoring. Accessed: 2024-11-18

    2016

  3. [11]

    C Bradford Barber, David P Dobkin, and Hannu Huhdanpaa. 1996. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS) 22, 4 (1996), 469–483

    1996

  4. [12]

    Sofia Belikovetsky, Mark Yampolskiy, Jinghui Toh, Jacob Gatlin, and Yuval Elovici

  5. [13]

    BMW Group. 2020. Additive Manufacturing at the BMW Group: 3D printing in series production. https://www.bmwgroup.com/en/news/general/2020/additive- manufacturing.html Accessed: 2024-05-30

    2020

  6. [14]

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. 2020. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)

    2020 arXiv

  7. [15]

    Gary R Bradski. 1998. Computer vision face tracking for use in a perceptual user interface. Intel technology journal 2 (1998)

    1998

  8. [16]

    Karl Bringmann, Nick Fischer, Ivor van der Hoog, Evangelos Kipouridis, Tomasz Kociumaka, and Eva Rotenberg. 2023. Dynamic Dynamic Time Warping. arXiv:2310.18128 [cs.CG] https://arxiv.org/abs/2310.18128

    2023 arXiv

  9. [17]

    Douglas Brion. 2024. OctoRest: Python client library for OctoPrint REST API. https://github.com/dougbrion/OctoRest

    2024

  10. [18]

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexan- der Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision . Springer, 213–229

    2020

  11. [19]

    Oliveira, Heitor Murilo Gomes, and André Grégio

    Fabrício Ceschin, Marcus Botacin, Albert Bifet, Bernhard Pfahringer, Luiz S. Oliveira, Heitor Murilo Gomes, and André Grégio. 2024. Machine Learning (In) CCS ’25, October 13–17, 2025, Taipei, Taiwan Twisha et al. Security: A Stream of Problems.Digital Threats 5, 1, Article 9 (...

    2024 doi

  12. [20]

    Fei Chen, Yuxi Luo, Nektarios Georgios Tsoutsos, Michail Maniatakos, Khaled Shahin, and Nikhil Gupta. 2019. Embedding tracking codes in additive manu- factured parts for product authentication. Advanced Engineering Materials 21, 4 (2019), 1800495

    2019

  13. [21]

    Yuming Chen, Xinbin Yuan, Ruiqi Wu, Jiabao Wang, Qibin Hou, and Ming-Ming Cheng. 2023. YOLO-MS: rethinking multi-scale representation learning for real- time object detection. arXiv preprint arXiv:2308.05480 (2023)

    2023 arXiv

  14. [22]

    Sujit Rokka Chhetri and Mohammad Abdullah Al Faruque. 2017. Side channels of cyber-physical systems: Case study in additive manufacturing. IEEE Design & Test 34, 4 (2017), 18–25

    2017

  15. [23]

    Sujit Rokka Chhetri, Arquimedes Canedo, and Mohammad Abdullah Al Faruque

  16. [24]

    CISA. 2024. Critical Infrastructure Sectors. https://www.cisa.gov/topics/critical- infrastructure-security-and-resilience/critical-infrastructure-sectors. Accessed: 2024-11-18

    2024

  17. [25]

    ACM Trans

    Confidentiality Breach Through Acoustic Side-Channel in Cyber-Physical Additive Manufacturing Systems. ACM Trans. Cyber-Phys. Syst. 2, 1, Article 3 (dec 2017), 25 pages. doi:10.1145/3078622

    2017 doi

  18. [26]

    Mark De Berg, Otfried Cheong, Olivier Devillers, Marc Van Kreveld, and Monique Teillaud. 1998. Computing the maximum overlap of two convex polygons under translations. Theory of computing systems 31, 5 (1998), 613–628

    1998

  19. [27]

    Dorin Comaniciu, Visvanathan Ramesh, and Peter Meer. 2000. Real-time tracking of non-rigid objects using mean shift. InProceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662) , Vol. 2. IEEE, 142– 149

    2000

  20. [28]

    M. A. Al Faruque, S. R. Chhetri, S. Faezi, and A. Canedo. 2016. Forensics of thermal side-channel in additive manufacturing systems. In CECS Technical Report No.16-01

    2016

  21. [29]

    Q. Do, B. Martini, and K. R. Choo. 2016. A data exfiltration and remote exploitation attack on consumer 3D printers. IEEE Transactions on Information Forensics and Security 11, 10 (10 2016), 2174–2186

    2016

  22. [30]

    Ford Media Center. 2023. Ford Opens New 3D Printing Centre to Support Production of its Future Vehicles. https://media.ford.com/content/fordmedia/ feu/en/news/2023/02/08/ford-opens-new-3d-printing-centre-to-support- production-of-its-f.html Accessed: 2024-05-30

    2023

  23. [31]

    Marlin Firmware. [n. d.]. M300 - Play Tone. https://marlinfw.org/docs/gcode/ M300.html. Accessed: 2025-06-23

    2025

  24. [32]

    Jacob Gatlin, Sofia Belikovetsky, Yuval Elovici, Anthony Skjellum, Joshua Lubell, Paul Witherell, and Mark Yampolskiy. 2021. Encryption is Futile: Reconstructing 3D-Printed Models Using the Power Side-Channel. In Proceedings of the 24th International Symposium on Research in A...

    2021

  25. [33]

    Yang Gao, Borui Li, Wei Wang, Wenyao Xu, Chi Zhou, and Zhanpeng Jin. 2018. Watching and safeguarding your 3D printer: Online process monitoring against cyber-physical attacks. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 2, 3 (2018), 1–27

    2018

  26. [34]

    GlobeNewswire. 2024. Global Additive Manufacturing Market Size To Worth USD 143.3 Billion By 2033: CAGR Of 21.78%. GlobeNewswire (2024). https://www.globenewswire.com/en/news-release/2024/04/19/2865997/0/en/ Global-Additive-Manufacturing-Market-Size-To-Worth-USD-143-3-Billion-...

    2024

  27. [35]

    General Motors. 2022. GM to Invest $81 Million to Build the Cadillac CELESTIQ at GM’s Global Technical Center. https://news.gm.com/newsroom.detail.html/ Pages/news/us/en/2022/jun/0615-celestiq-gm.html Accessed: 2024-05-30

    2022

  28. [36]

    Sinan Gültekin, Ahmet Ural, and Ulas Yaman. 2019. Embedding QR codes on the interior surfaces of FFF fabricated parts. Procedia Manufacturing 39 (2019), 519–525

    2019

  29. [37]

    GrabCAD. 2015. GrabCAD Library: Open Source Engineering Designs. https: //grabcad.com/library. Accessed: 2024-11-18

    2015

  30. [38]

    Keygen. 2024. Eric Van Albert. https://keygen.co/

    2024

  31. [39]

    Berthold KP Horn and Brian G Schunck. 1981. Determining optical flow.Artificial intelligence 17, 1-3 (1981), 185–203

    1981

  32. [40]

    Killick, P

    R. Killick, P. Fearnhead, and I. A. Eckley. 2012. Optimal Detection of Changepoints With a Linear Computational Cost. J. Amer. Statist. Assoc. 107, 500 (Oct. 2012), 1590–1598. doi:10.1080/01621459.2012.737745

    2012

  33. [41]

    Ryosuke Kikuchi, Sora Yoshikawa, Pradeep Kumar Jayaraman, Jianmin Zheng, and Takashi Maekawa. 2018. Embedding QR codes onto B-spline surfaces for 3D printing. Computer-Aided Design 102 (2018), 215–223

    2018

  34. [42]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. 2017. ImageNet classification with deep convolutional neural networks. Commun. ACM 60, 6 (5 2017), 84–90. https://doi.org/10.1145/3065386

    2017 doi

  35. [43]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Opti- mization. arXiv:1412.6980 [cs.LG]

    2017 arXiv

  36. [44]

    Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al . 2022. YOLOv6: A single- stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)

    2022 arXiv

  37. [45]

    Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. 2018. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE conference on computer vision and pattern recognition . 8971–8980

    2018

  38. [46]

    Jimmy Liranzo and Thaier Hayajneh. 2017. Security and privacy issues affecting cloud-based IP camera. In 2017 IEEE 8th Annual Ubiquitous Computing, Electronics and Mobile Communication Conference (UEMCON) . IEEE, 458–465

    2017

  39. [47]

    Sizhuang Liang, Saman Zonouz, and Raheem Beyah. 2022. Hiding my real self! protecting intellectual property in additive manufacturing systems against optical side-channel attacks. In Proceedings 2022 Network and Distributed System Security Symposium. Internet Society

    2022

  40. [48]

    Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic Gradient Descent with Warm Restarts. arXiv:1608.03983 [cs.LG]

    2017 arXiv

  41. [49]

    Marcus Littwin. 2024. Repetier.com. https://www.repetier.com/

    2024

  42. [50]

    Terrence Mativo, Christian Fritz, and Ismet Fidan. 2018. Cyber acoustic analy- sis of additively manufactured objects. The International Journal of Advanced Manufacturing Technology 96, 1 (4 2018), 581–586. doi:10.1007/s00170-018-1603-z

    2018 doi

  43. [51]

    Upal Mahbub, Hafiz Imtiaz, and Md Atiqur Rahman Ahad. 2011. An optical flow based approach for action recognition. In 14th international conference on computer and information technology (ICCIT 2011) . IEEE, 646–651

    2011

  44. [52]

    Meinard Müller. 2020. Dynamic Time Warping (DTW) for Music Pro- cessing. https://www.audiolabs-erlangen.de/resources/MIR/FMP/C7/C7S2_ SubsequenceDTW.html Accessed: 2024-11-12

    2020

  45. [53]

    Adriano Moreira and Maribel Yasmina Santos. 2007. Concave hull: A k-nearest neighbours approach for the computation of the region occupied by a set of points. (2007)

    2007

  46. [54]

    Python.org. 2024. Python Data Structures. https://docs.python.org/3/tutorial/ datastructures.html#nested-list-comprehensions. Accessed: 2024-11-18

    2024

  47. [55]

    OctoPrint. 2024. The snappy web interface for your 3D printer. https://octoprint. org/

    2024

  48. [56]

    SHENZHEN GETECH TECHNOLOGY CO. LTD. 2020. Geeetech A20T 3D Printer User Manual (V1.10). https://www.geeetech.com/download.html?download_id= 45

    2020

  49. [57]

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. 2016. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition . 779–788

    2016

  50. [58]

    Logan D Sturm, Christopher B Williams, Jamie A Camelio, Jules White, and Robert Parker. 2017. Cyber-physical vulnerabilities in additive manufacturing systems: A case study attack on the. STL file with human subjects. Journal of Manufacturing Systems 44 (2017), 154–164

    2017

  51. [59]

    Cheng Song, Fan Lin, Zhigang Ba, Kui Ren, Chen Zhou, and Wenyuan Xu. 2016. My smartphone knows what you print: Exploring smartphone-based side-channel attacks against 3D printers. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security (CCS ’16...

    2016

  52. [60]

    Thingiverse. 2024. Digital Designs for Physical Objects. https://www.thingiverse. com/

    2024

  53. [61]

    The Open Locksport Guide. 2006. Understanding Key Bitting. http:// lockpickernetwork.wikidot.com/understanding-key-bitting

    2006

  54. [62]

    UltiMaker. 2024. UltiMaker Cura. https://ultimaker.com/software/ultimaker- cura/

    2024

  55. [63]

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. 2022. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. arXiv:2203.12602 [cs.CV]

    2022 arXiv

  56. [64]

    Williams

    Linda D. Williams. 2015. Additive Manufacturing or 3d Scanning and Printing. In Manufacturing Engineering Handbook. McGraw-Hill, New York

    2015

  57. [65]

    Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. 2019. Fast online object tracking and segmentation: A unifying approach. InProceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition . 1328– 1338

    2019

  58. [66]

    Wong and A

    K.V. Wong and A. Hernandez. 2012. A review of additive manufacturing. ISRN Mech. Eng. (2012). doi:10.5402/2012/208760

    2012 doi

  59. [67]

    Nicolai Wojke, Alex Bewley, and Dietrich Paulus. 2017. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP). IEEE, 3645–3649

    2017

  60. [68]

    Zhiyuan Yu, Yuanhaur Chang, Shixuan Zhai, Nicholas Deily, Tao Ju, XiaoFeng Wang, Uday Jammalamadaka, and Ning Zhang. 2023. XCheck: Verifying In- tegrity of 3D Printed Patient-Specific Devices via Computing Tomography. In 32nd USENIX Security Symposium (USENIX Security 23) . US...

    2023

  61. [69]

    Yanqing Yao, Gong Cheng, Guangxing Wang, Shengyang Li, Peicheng Zhou, Xingxing Xie, and Junwei Han. 2023. On Improving Bounding Box Representa- tions for Oriented Object Detection. IEEE Transactions on Geoscience and Remote Sensing 61 (2023), 1–11. doi:10.1109/TGRS.2022.3231340

    2023

  62. [70]

    Zhilu Zhang and Mert R. Sabuncu. 2018. Generalized cross entropy loss for training deep neural networks with noisy labels. In Proceedings of the 32nd International Conference on Neural Information Processing Systems (Montréal, Canada) (NIPS’18). Curran Associates Inc., Red Hoo...

    2018

  63. [71]

    Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. 2021. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision 129 (2021), 3069–3087. One Video to Steal Them All: 3D-Printing IP Theft ...

    2021

  64. [2017]

    In 11th USENIX workshop on offensive technologies (WOOT 17)

    dr0wned–{Cyber-Physical} attack with additive manufacturing. In 11th USENIX workshop on offensive technologies (WOOT 17)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.