Writing study: Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Lei Wang, Wenxiang Diao, Andrew Busch, Jun Zhou, Yongsheng Gao · Official ICML 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Architecture & representation (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

33 physical pages

Before references / appendix

≈ 9.25 page-equivalents

Main figures / tables

6 / 8

Attached appendix figures / tables

0 / 21

Attached appendix span

≈ 19.91 page-equivalents

Before references / appendix: ≈ 28.03%References / intervening material: ≈ 11.64%Attached appendix: ≈ 60.33%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1. Introduction	1	≈ 1.3 pp.	3.94%
2. Related Work	2	≈ 0.62 pp.	1.88%
3. Method	3	≈ 2.1 pp.	6.36%
4. Experiment	5	≈ 4.5 pp.	13.64%
5. Conclusion	9	≈ 0.18 pp.	0.55%
Impact Statement	10	≈ 0.05 pp.	0.15%
ACKNOWLEDGMENTS	10	≈ 0.16 pp.	0.48%
References	10	≈ 3.84 pp.	11.64%
A. Where Privacy Can Be Protected?	14	≈ 1.07 pp.	3.24%
B. Why Privacy is Protected?	15	≈ 0.42 pp.	1.27%
C. Terminology and Conceptual Clarifications	15	≈ 0.68 pp.	2.06%
D. Connections to Related Work	16	≈ 0.92 pp.	2.79%
E. Theoretical Motivation: Sensitive Attributes in Linearly Separable Subspaces	17	≈ 0.98 pp.	2.97%
F. Beyond a Linear Layer: How OPL/G-OPL Differ from Standard FC Layers	18	≈ 0.57 pp.	1.73%
G. Algorithmic Implementation of OPL and G-OPL	18	≈ 1.35 pp.	4.09%
H. Practical Considerations for Layer Placement	20	≈ 1 pp.	3.03%
I. Why Stack Multiple OPL/G-OPL Layers?	21	≈ 0.55 pp.	1.67%
J. Discussions on Privacy-Aware Evaluation Metrics	21	≈ 1.14 pp.	3.45%
K. Detailed Setups and Configurations	22	≈ 0.93 pp.	2.82%
L. Additional Results and Evaluations	23	≈ 4.86 pp.	14.73%
M. Layer Placement: Practical Guidelines and Empirical Insights	28	≈ 4.12 pp.	12.48%
N. Limitations	32	≈ 0.48 pp.	1.45%
O. LLM Usage Declaration	33	≈ 0.82 pp.	2.48%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 2
• Figure 3 · p. 5
• Table 1 · p. 5
• Table 2 · p. 6
• Table 3 · p. 6
• Figure 4 · p. 6
• Figure 5 · p. 7
• Table 4 · p. 8
• Figure 6 · p. 8
• Table 5 · p. 8
• Table 6 · p. 9
• Table 7 · p. 9
• Table 8 · p. 9

Attached appendix

• Table 9 · p. 22
• Table 10 · p. 23
• Table 11 · p. 24
• Table 12 · p. 24
• Table 13 · p. 25
• Table 14 · p. 25
• Table 15 · p. 26
• Table 16 · p. 27
• Table 17 · p. 27
• Table 18 · p. 27
• Table 19 · p. 28
• Table 20 · p. 28
• Table 21 · p. 28
• Table 22 · p. 29
• Table 23 · p. 29
• Table 24 · p. 29
• Table 25 · p. 29
• Table 26 · p. 30
• Table 27 · p. 30
• Table 28 · p. 31
• Table 29 · p. 31

Argument and organization

1. 1 Introduction; 2 Related Work

   Frames anomaly detection as a utility and representation-privacy problem.

2. 3 Method

   Defines the learned projection, weakly guided sensitive subspace and diagnostic metrics.

3. 4 Experiment; 5 Conclusion

   Studies module placement, rank, utility, privacy probes, cost and multiple attribute signals.

Problem, claim and contribution

Privacy is operationalized as suppression of sensitive information in intermediate features, rather than a formal privacy budget or raw-data anonymization guarantee.

Abstract, p.1; metrics, pp.4-5 · Main paper

Argument and section order

The projection equation is followed by its guided loss and metric definitions before comparisons, making the intervention and its intended measurement explicit.

Section 3, pp.3-5 · Main paper

Figures, captions and visual explanation

The pipeline distinguishes guided and unguided modules by color; the dense results table groups utility and privacy columns by dataset and marks unavailable face detections rather than imputing values.

Figure 2, p.2; Table 4, p.8 · Main paper

Tables, comparisons and evidence

The paper adds an ArcFace retrieval probe beside face-presence and subspace metrics, and reports a utility decline when suppressing task-relevant gait.

Tables 4 and 7-8, pp.8-9 · Main paper

Appendix and reproducibility

Deployment discussion distinguishes client-side filtering from server-side filtering that still exposes raw video; another appendix clarifies which metrics require a projection module.

Appendix H, p.20; Appendix J, p.22 · Main paper

Clarity and writing conventions

Application exceptions and placement sensitivity qualify broad privacy language: face suppression is optional when identity is necessary to define the anomaly, and unseen domains require further validation.

Appendix A, p.14; Appendix N, pp.32-33 · Main paper

What to reuse—and what not to infer

Transferable lesson: Define the privacy quantity being measured and separate representation diagnostics, attacker probes and deployment data flow before making a broad protection claim.

Boundary: Low linear-probe accuracy is not a guarantee against every re-identification attack, and feature filtering does not protect raw video already sent to a server. Module placement and the relevance of the suppressed attribute remain task-dependent.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan of the main argument through physical p.9, with focused physical-page text inspection on 1, 2, 3, 4, 5, 6, 7, 8, 9. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 14, 16, 20, 22, 32, 33. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

2, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 33 pages