A Category-Selective Encryption Framework for Privacy Protection of UAV Images
Ting Yan, Minghui Shang, Haolang Sun, Xu Wang, Yamin Han, Mingze He
Abstract
Imagery captured by unmanned aerial vehicle (UAV) systems increasingly raises privacy concerns, as UAV aerial surveillance routinely captures sensitive objects such as pedestrians and vehicles alongside ordinary scene content. Existing privacy-preserving approaches typically employ class-agnostic saliency detection to indiscriminately obfuscate all visually salient regions, leading to unnecessary information loss and hindering downstream scene analysis. This paper proposes a category-selective encryption framework for privacy-preserving UAV surveillance. The core advance over prior class-agnostic approaches is semantic-level selectivity: the system first recognizes what is present in the scene via an 80-class detector, then encrypts only user-specified sensitive categories while leaving non-sensitive objects and background intact for downstream analytics. The framework supports two complementary encryption schemes: Thumbnail-Preserving Encryption (TPE), which applies adaptive pixelation with a learnable blend parameter and recovers content through a jointly trained Enhanced U-Net, and XOR stream cipher encryption, which offers parameter-free, mathematically exact reversibility. The architectural design, composite loss formulation, and joint training methodology are presented in detail, establishing a foundation for practical privacy-preserving UAV surveillance that balances protection strength and scene utility.
Review reports
xPeerd (AI review)
AI Review (xPeerd — DAReview)
Summary
This manuscript, titled A Category-Selective Encryption Framework for Privacy Protection of UAV Images, proposes a UAV-image privacy framework that detects objects with YOLO11n and then selectively encrypts only user-specified sensitive categories rather than the whole image. It offers two branches: a TPE-based branch with adaptive pixelation plus an Enhanced U-Net for approximate restoration, and an XOR stream-cipher branch for exact reversible recovery. The paper’s main strength is the clear practical motivation: preserving background content for downstream scene analysis while obscuring sensitive targets. The architecture and intended workflow are described coherently, and the comparative framing against full-image encryption is sensible.
Its main weaknesses concern validation depth, security rigor, and reproducibility. The experimental section emphasizes PSNR/SSIM and latency, but provides limited evidence for cryptographic security, object-detection robustness under realistic deployment conditions, and restoration reliability across broader datasets or threat models. Some claims are stronger than the evidence presented, especially where the manuscript infers practical suitability for real-time UAV transmission and privacy protection without a fuller adversarial or systems-level evaluation. There are also several writing and formatting issues, plus a few technically imprecise or overstated sentences. Overall, the manuscript is promising as an engineering prototype, but it presently needs substantial strengthening in methodological justification, evaluation breadth, and presentation quality before it can be considered mature.
Potential Major Revisions
- Security evaluation is underdeveloped relative to the manuscript’s central privacy claim.
The paper repeatedly frames the method as an image privacy protection or encryption framework, yet the evaluation is dominated by reconstruction quality and latency rather than security analysis. For example, the abstract claims the work establishes “a foundation for practical privacy-preserving UAV surveillance that balances protection strength and scene utility” , and the conclusion states that it “simultaneously delivers visual privacy protection while preserving basic image content” and is “suitable for real-time image transmission scenarios of UAVs” . However, the evidence shown in the retrieved sections mainly consists of PSNR/SSIM comparisons and qualitative visual examples .
That is not enough for a paper whose core contribution is privacy-preserving encryption. In particular, the XOR branch is presented as “parameter-free, mathematically exact reversibility” and as using “the user-defined key string” mapped through SHA-256 into a seed for PCG-64 pseudo-random noise generation . Yet no cryptographic threat model is stated, and no analysis is provided for key sensitivity, resistance to known-plaintext/chosen-plaintext attacks, security of the PRNG choice for cryptographic use, or leakage through preserved background. A privacy paper should distinguish visual obfuscation from cryptographic security. As written, the manuscript risks overstating the XOR path as strong encryption without demonstrating that the pseudo-random stream generator and masking strategy meet accepted security expectations.
The TPE branch has a parallel issue. The manuscript states that “When α = 1, the region is fully pixelated with the strongest privacy protection” . This is an assertion, not a demonstrated result. Pixelation strength is not equivalent to privacy strength unless the paper evaluates re-identification resistance, face/license-plate leakage, inversion risk, or attack-based recovery from mosaiced regions. The authors should substantially revise the paper to include a clearly defined threat model and security metrics appropriate to each branch.
- The manuscript conflates image restoration quality with privacy-preserving success.
The experimental section interprets high SSIM/PSNR after decryption as strong evidence of framework quality. For example, “XOR encryption achieves an SSIM of 1 and infinite PSNR” and “demonstrates that XOR enables lossless pixel-level reconstruction” . While this is valid for reversibility, it says nothing about the privacy strength of the encrypted image itself. Likewise, for TPE, the paper emphasizes that “the textures and contours of targets can be highly restored” , but that is a restoration outcome, not a privacy metric.
The paper should separate three questions that are currently blended together:
- how well sensitive content is concealed before decryption,
- how well original content is recovered after decryption, and
- how well non-sensitive background remains usable for downstream tasks.
At present, the manuscript mainly addresses item 2 and partly item 3. Item 1 is insufficiently evaluated. This is a major conceptual gap because the paper’s title and motivation center on privacy protection. Without a stronger evaluation of concealment effectiveness, claims of privacy protection remain only partially supported.
- Detection dependence is a critical vulnerability but is not rigorously examined.
The framework encrypts only regions detected by YOLO11n. The paper states that “the lightweight YOLO11n detector can accurately localize sensitive objects and generate compact bounding boxes” even in “complex aerial scenarios” . It also claims the object detector “can flexibly adapt to encryption requirements of various scenarios and distinguish sensitive and non-sensitive targets” . However, these are consequential assumptions: if detection fails, privacy fails.
This dependency should be treated as a central limitation and experimentally stress-tested. The current text does not provide enough evidence about missed detections, false negatives on privacy-sensitive classes, small-object behavior, occlusion robustness, night/blur conditions, or domain shift between COCO pretraining and aerial imagery deployment. The manuscript notes the detector is “pre-trained on the MS COCO dataset” and recognizes “80 common object categories” , but the paper does not explain whether it was fine-tuned on VisDrone for privacy-sensitive aerial classes, how detection performance transfers, or how privacy risk is handled when sensitive targets are not detected.
Because missed detections directly expose private content, this omission is major. The authors should report detection performance for the sensitive categories actually used in encryption, discuss failure cases, and quantify the privacy consequences of detector errors.
- The experimental design appears too narrow for the claimed generality.
The manuscript states: “All experiments are carried out on the public UAV aerial dataset VisDrone” . A single-dataset evaluation may be acceptable for an initial prototype, but the paper makes broader claims about practical UAV surveillance use . Since the method depends on detection, scene composition, target scale, and category selection, broader validation is needed.
There is also limited detail in the retrieved text about train/validation/test protocol, statistical variability, ablation design, and whether results are averaged over multiple runs or scenes. The training description indicates that “we freeze the YOLO11 detection backbone” and optimize “the learnable blending coefficient α of TPE and all weights of the Enhanced U-Net restoration network” . But there is not enough information here about training duration, optimization settings, uncertainty intervals, or sensitivity to α regularization. For a method paper, that weakens reproducibility and confidence in the reported gains.
- Several comparative claims are stronger than the presented evidence.
The paper writes that the proposed scheme “outperforms saliency-based encryption schemes without category discrimination” and that “the proposed schemes achieve better reconstruction performance than all comparative baselines within their respective technical branches” . These are broad superiority claims. Yet the visible comparison appears limited to HF-TPE and TPE-GAN in Table III and visual examples in Fig. 3 . The paper may be directionally correct, but it should be more restrained unless the benchmark space is comprehensive and methodologically balanced.
For example, HF-TPE is full-image encryption, while Proposed-XOR is region-selective. This is not a like-for-like comparison in terms of objective or operating assumptions. Similarly, comparing a deep-learning branch and a cryptography-style branch only via PSNR/SSIM and latency simplifies a more complex tradeoff space. The authors should either soften these claims or provide a fuller benchmarking and fairness discussion.
- Reproducibility of the TPE branch is incomplete and the methodological rationale needs sharpening.
The paper introduces the joint loss “Ltotal = Lrec + λαLα” with “λα ... set to 0.6” and defines “Lα = 1 − α” to prevent the model from weakening encryption merely to improve reconstruction . This is interesting, but under-justified. Why 0.6? Was it tuned? How sensitive are results to this value? Is α global, per-image, per-box, or learned under a more granular parameterization? The text says “learnable blending coefficient α” in the singular , which suggests one global scalar, but that design choice should be justified because privacy requirements may vary by object scale, class, or context.
Likewise, the adaptive block size is defined by “k = max{kmin, floor(min(h,w)/d)}” with “kmin = 6” and “d = 8” . This is plausible, but it reads heuristic. A stronger manuscript would justify these choices empirically and show ablations on block-size policy, α-regularization, detector thresholds, and restoration architecture.
- The manuscript shows some signs of overstatement and internal tension in its claims.
One recurring tension is between preserving utility and claiming privacy/security. The paper states that non-sensitive pixels are “completely retaining the semantic information of non-sensitive objects and backgrounds” and that only targeted classes are encrypted for downstream analytics . That is a design advantage, but it may also leak contextual information relevant to sensitive-object inference. For example, if a scene preserves trajectories, interactions, surrounding objects, or visible body parts outside a bounding box, privacy protection may be weaker than implied. This is especially important when bounding boxes are compact rather than segmentation-based. The manuscript should discuss privacy leakage through preserved context instead of assuming that box-level obfuscation is sufficient.
A second tension concerns the sentence “If Cenc = ∅, all detected objects will be encrypted” . This means the system still does not encrypt the whole image, only all detected objects. That should be stated carefully, because the distinction matters for privacy coverage and for comparisons against full-image methods.
- Paper-mill or fabrication warning: low to moderate concern, but some indicators merit scrutiny.
I do not see strong evidence of fabrication from the retrieved text. The problem, architecture, and formulas are broadly coherent. However, several features warrant caution: frequent use of inflated evaluative wording such as “drastically,” “highly,” “favorable,” and “outstanding” ; claims of practical suitability without broader systems validation; and a somewhat thin security assessment for a paper centered on encryption. These do not prove fraud, but they do suggest that the manuscript may overmarket an engineering proof-of-concept. I would therefore advise careful human verification of experimental completeness, implementation details, and the correctness of security terminology before publication.
Potential Minor Revisions
- Wording, grammar, and style issues
Several sentences are understandable but not polished to publication level.
- The abstract says “The core advance over prior class-agnostic approaches is semantic-level selectivity” . This is acceptable, but the surrounding prose is dense and could be simplified for readability.
- In the introduction, “personal privacy will be exposed, which will lead to privacy breaches” is redundant and stylistically awkward .
- “The proposed TPE scheme is the deep learning-based encryption method designed in the proposed” is grammatically incorrect and should be rewritten .
- “This paper designs supporting decryption pipelines” is slightly unnatural; “develops” or “implements” would read better .
- “The whole noise map is generated at one time rather than box by box” is colloquial and should be revised to “generated in a single pass” or similar .
- “The drastically reduced number of encrypted pixels significantly lowers the computational burden” is stylistically repetitive because both “drastically” and “significantly” intensify the same claim .
- “Future work will further optimize the network inference speed and encryption security strength” is understandable but awkwardly phrased; “improve inference speed and strengthen security evaluation/design” would be more precise .
- Terminological precision
- The manuscript alternates between privacy protection, encryption, shielding, obfuscation, and restoration in ways that may confuse readers about which guarantees are cryptographic and which are visual. This is especially important because the TPE branch is not equivalent to conventional encryption in the strict cryptographic sense. Relevant passages include the abstract’s description of TPE and XOR as “two complementary encryption schemes” and the claim that “When α = 1, the region is fully pixelated with the strongest privacy protection” . The terminology should be made more precise throughout.
- Formatting issues
- The PDF text suggests some line-break and spacing artifacts, such as “stream ci- pher encryption” in the contribution list .
- Section and figure formatting appears uneven in the extracted text, especially around equations, table headings, and figure captions .
- Capitalization style is inconsistent in places, for example “person & car” in Table II versus broader category wording in the prose .
- The authors should carefully check table titles, figure captions, and page layout for overflow and alignment problems.
- Missing reporting details that should be added even if not central enough to count as major conceptual flaws
- Explicit training hyperparameters for the Enhanced U-Net branch.
- Hardware platform details for the latency numbers in Table IV, since “∼0.5ms,” “∼1ms,” and “∼30ms” are hardware-dependent .
- Clarification of whether page-level, image-level, or object-level averages are used in reporting PSNR/SSIM.
- Confidence intervals or standard deviations for quantitative results.
- Detector fine-tuning protocol and performance for the specific sensitive classes.
- AI content analysis for this post-2021 manuscript
Estimated percentage of AI-generated content: 28%–42%.
Basis for this estimate:
- The manuscript is coherent and technically organized, which weighs against heavy uncontrolled AI generation.
- However, several passages use formulaic promotional academic phrasing and generic transition patterns typical of AI-assisted drafting, for example: “establishing a foundation for practical privacy-preserving UAV surveillance that balances protection strength and scene utility” ; “drastically reduces redundant computation” ; “favorable balance between privacy protection and image usability” .
- There are also recurring broad claims that sound rhetorically polished but empirically under-specified.
Highlighted AI-detected sections
- Abstract, especially the compressed contribution-summary style and broad framing language .
- Introduction passages that repeatedly contrast “privacy protection” and “scene usability” in generalized language .
- Comparative-analysis prose using repeated evaluative descriptors such as “favorable,” “drastically,” and “outstanding” .
Assessed epistemic impact with respect to estimated and claimed contributions:
- If AI assistance was used only for language polishing, the epistemic impact is low.
- If AI assistance contributed to technical exposition, then the main risk is not fabrication in the retrieved sections but imprecision, overclaiming, and insufficient distinction between visual obfuscation and cryptographic security.
- The manuscript’s core contribution remains plausible, but any AI-assisted drafting would heighten the need for human verification of security claims, methodological rationale, and benchmark fairness.
Recommendations
-
Add a formal threat model. Specify what adversary is assumed for each branch, what information is exposed, what constitutes a privacy breach, and whether the goal is cryptographic confidentiality, visual anonymization, or both.
-
Expand privacy/security evaluation. For XOR, include key sensitivity, randomness/security justification for the keystream generation method, attack-oriented analysis, and discussion of whether PCG-64 is appropriate for security-critical encryption. For TPE, include privacy leakage or re-identification tests, not just reconstruction metrics.
-
Report detection performance as a first-class privacy metric. Because the system relies on detection before encryption, provide precision/recall or mAP for privacy-sensitive classes in aerial conditions, analyze false negatives explicitly, and discuss failure handling.
-
Strengthen reproducibility. Provide complete implementation details, training setup, hyperparameters, hardware configuration, and ablation studies for α, λα, block-size policy, detector thresholds, and restoration architecture choices.
-
Moderate or qualify broad claims. Statements about outperforming prior work, practical suitability, and strongest privacy protection should be rewritten more cautiously unless supported by broader evidence.
-
Discuss limitations candidly. The manuscript should explicitly acknowledge that box-level selective encryption may leak contextual cues, that detector misses expose privacy-sensitive content, and that VisDrone-only evaluation limits generalization.
-
Revise language and presentation. A full editorial pass is needed to correct awkward phrases, reduce promotional wording, improve grammatical consistency, and polish tables/figures.
Overall recommendation
The manuscript contains a relevant and potentially useful systems idea: combining category-aware object detection with selective privacy processing for UAV imagery. However, in its current form, I would regard it as not yet ready for acceptance without major revision. The largest issues are insufficient security validation, heavy dependence on unquantified detector reliability, limited experimental breadth, and recurring overstatement relative to the presented evidence.
XSci (AI review)
Summary
The manuscript addresses privacy in UAV aerial imagery, where blanket encryption or class-agnostic saliency-based obfuscation destroys scene content that downstream analytics need. The authors propose a closed-loop pipeline in which a frozen COCO-pretrained YOLO11n detector labels all objects, a user-defined sensitive-category set filters the detections (Eq. 4), and only the selected boxes go to one of two interchangeable branches: thumbnail-preserving encryption (TPE), which pixelates each box and blends it with the original through a learnable coefficient α (Eq. 6) before an Enhanced U-Net restores it, and an XOR stream cipher keyed through SHA-256 and PCG-64. On VisDrone the TPE branch reports SSIM 0.9792 / PSNR 38.48 dB for people only and 0.9734 / 35.26 dB with vehicles added; the XOR branch reports SSIM 1 and infinite PSNR.
Contribution and Significance
The core contribution is an engineering one: inserting a user-specified category filter between an off-the-shelf multi-class detector and an encryption primitive, so that the choice of what to protect becomes a runtime parameter rather than a property of the algorithm. The accompanying training recipe — a mask-restricted L1 objective (Eq. 9) paired with a regularizer that pushes α toward 1 (Eq. 10) — is the most thoughtful part of the design, because the authors correctly anticipated that a reconstruction-only loss would collapse α to zero and remove the encryption altogether.
The novelty relative to the closest prior work is smaller than the manuscript claims. Detection-driven selective encryption with user-chosen targets is established: Asgari-Chenaghlu et al., "Cy: Chaotic YOLO for user intended image encryption and sharing in social media," Information Sciences 542:212–227, 2021, already encrypts user-selected YOLO-detected regions, and Song, Fu, Zheng, Cao, Tie and Sham, "Protection of image ROI using chaos-based encryption and DCNN-based object detection," Neural Computing and Applications 34(7):5743–5756, 2022, builds a multi-object ROI cipher on YOLOv4 and analyses the box-edge leakage this manuscript ignores. Neither is cited, and moving to YOLO11 and from a chaotic map to XOR is not in itself an advance. Section I's claim that prior UAV schemes "fail to achieve discriminative selective encryption" is also contradicted by reference [10], which encrypts detected face regions selectively.
Strengths
- The pipeline is specified end to end without ambiguity (Fig. 1, Eq. 2), and the two branches are presented honestly as having genuinely different reversibility properties.
- The pairing of Eqs. (9) and (10) is a correct and non-obvious design decision: the authors identified the degenerate solution mask-restricted L1 alone would produce and added the counter-pressure that prevents it.
- The scale-adaptive block size of Eq. (5) responds to a real difficulty of VisDrone imagery, where object sizes span more than an order of magnitude within a frame.
- Table II reports the utility cost of widening the sensitive set rather than concealing it, restoration falling from 0.9792 / 38.48 dB to 0.9734 / 35.26 dB as vehicles are added.
Recommendation
Reject for IEEE MIPR. The central claim is not supported and, for the TPE branch, cannot be supported within the manuscript's approach: that branch has no key, and the XOR branch's keystream is a function of the key alone, so it repeats across every image. Neither quantity the framework trades off — privacy strength and downstream utility — is measured anywhere in Section IV, and Table IV contradicts the text it summarises. The idea is worth pursuing; the evidence here does not yet establish it.
Major Concerns
-
Neither branch is secure as specified, and no security evaluation is reported (Sections III.B.1, III.B.2, IV). The TPE branch has no key: restoration depends only on the Enhanced U-Net weights and α, so any party able to train an equivalent network on pixelated aerial crops — which the authors show is easy — recovers the protected content. This is the failure mode documented by McPherson, Shokri and Shmatikov, "Defeating Image Obfuscation with Deep Learning," arXiv:1609.00408, 2016; the reported 38.48 dB restoration is evidence for the attack, not against it.
The XOR branch fares little better. Section III.B.2 derives the full-resolution noise map from the key alone and states it "is generated at one time rather than box by box", so one keystream covers every image encrypted under that key; since the mask moves between frames, a pixel encrypted in one frame and left in the clear in the next discloses the keystream at that location, with no cryptanalysis required. PCG-64 is in any case not a cryptographic generator — its 128-bit state is recoverable from 192 bits of output (Bouillaguet, Martinez and Sauvage, IACR Trans. Symmetric Cryptology, 2020). Section IV reports none of the key-space, key-sensitivity, histogram, correlation, entropy or NPCR/UACI analyses that baselines [7]–[10], [15], [17] give; a method called encryption needs a threat model and an evaluation against it.
-
The two quantities the framework exists to trade off are never measured (Sections IV.A, IV.B; contribution 3 in Section I). Section IV.A claims the system "sustains stable object detection accuracy on encrypted images", yet no mAP, precision or recall figure appears anywhere; privacy is evidenced only by visual inspection of Fig. 2, with no recognition accuracy measured on encrypted regions. Protection is only as complete as YOLO11n's recall, and a COCO-pretrained detector on small overhead VisDrone targets is precisely where recall is weak — every missed pedestrian is an unprotected pedestrian, and that rate is never reported. The category vocabulary is also mismatched with the motivation: the abstract and Section I frame the problem around "human faces and vehicle license plates", neither among COCO's 80 classes, and Section III.A.3 calls license plates sensitive and vehicles "non-sensitive" while Section IV.A encrypts vehicles. Detection recall on sensitive classes, a downstream-task metric, and a privacy metric on protected regions would resolve this.
-
The reconstruction comparison in Tables II and III is not like-for-like (Sections IV.B, IV.C). PSNR and SSIM are computed "between the original aerial images and decrypted images", i.e. over the whole frame, yet only detection boxes are altered — a small fraction of pixels in aerial imagery — so the values are dominated by background that is bit-identical by construction and are not comparable with HF-TPE [15] or TPE-GAN [17], which encrypt the entire image. The in-mask PSNR/SSIM that would settle this is absent. "SSIM 1, PSNR inf" for Proposed-XOR is a definitional property of any involutive cipher, not a measured outcome, yet Section IV.C presents it as outperforming HF-TPE's 0.9996 / 55.40 dB; and whether the baseline numbers were re-run on VisDrone is not stated.
-
Section IV.D and Table IV contradict each other and the method section (Table IV, Section IV.D). The text gives XOR "an average latency of merely 0.5 ms" with "no trainable network parameters", TPE encryption "approximately 1 ms", and TPE decryption "50.29 million trainable parameters" and "around 30 ms"; Table IV assigns the opposite — 0.5 ms and N/A to the TPE rows, 1 ms, 50.29M and 30 ms to the XOR rows. The same paragraph calls α "a manually defined hyperparameter", contradicting Sections III.B.1 and III.D.1, where α is the learnable coefficient contribution 3 rests on. It also claims "both modules restrict all computational operations within sensitive bounding boxes", which cannot hold when Section III.C.1 feeds the whole encrypted image to the U-Net and Section III.B.2 generates a full-resolution noise map in one shot, so the efficiency argument for onboard deployment does not survive.
Minor Comments
- The Enhanced U-Net is described only as having "SE channel attention with a channel compression ratio r = 8" and 50.29M parameters, so it cannot be reproduced.
- Report the converged value of α and a sensitivity study for λ_α, fixed at 0.6 without justification.
- Excluding background pixels from the loss does not "guarantee consistency between the background area and the original image" (Section III.D.1), since unsupervised output is unconstrained; state whether the background is composited back.
- Give the experimental protocol: VisDrone split, image counts per split, and how early stopping was monitored.
- The rationale for Eq. (5) is inverted: k = max(6, ⌊min(h,w)/8⌋) makes block size proportional to object scale, so relative blur is constant, not stronger for larger targets.
- Rectangular encrypted regions still disclose the number, position and size of sensitive objects; discuss this residual leakage.
- Class-agnostic saliency encryption is the motivating contrast, yet no saliency baseline such as [11] is evaluated, and "U-2-Net", named in Section III.A.3, is absent from the reference list.
- Freezing the detector does not "preserve consistent object detection accuracy on encrypted images" as contribution 3 states, since detection runs on the plaintext image.
- In Eq. (7) the index i is undefined, and the mask is declared M ∈ {0,1}^{H×W} after Eq. (2) but {0,255}^{H×W} in Section III.B.2 without a stated conversion point.
- The latency figures exclude YOLO11n inference, which dominates the real pipeline, and the "unified experimental environment" is never specified.
- Discuss Asgari-Chenaghlu et al. (2021) and Song et al. (2022), cited in full above.
- Align the motivating examples with the classes the system can act on, since an 80-class COCO detector cannot localise faces or license plates.