Pith. sign in

REVIEW 4 major objections 6 minor 26 references

Hiding in Plain Sight: Query Obfuscation via Random Multilingual Searches

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Random multilingual queries can shift a search engine's inferred interests without changing displayed results.

desk verdict Genuinely interesting empirical setup, but the headline significance claims rest on autocorrelated, unreplicated data. read the letter →

arxiv 2506.04963 v1 pith:NVV5IHVE submitted 2025-06-05 cs.CR

classification cs.CR
keywords queryobfuscationuserprofilingsearchenginepersonalizationmultilingualrandomqueriesprivacy-enhancingtechnologySeznam.czdecoytrafficprofileoverwriting
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

The paper tries to establish that a user can hide their real interests from a search engine simply by having a client-side tool fire off random queries in a mix of languages between their genuine searches. It tests this on Seznam.cz, a Czech engine that openly shows the interest profile it infers for each account. Across four controlled experiments, the paper finds that the engine's displayed top results barely change, but the inferred interest profile shifts substantially, and that decoy traffic can even reshape a profile that has already formed. The practical payoff, if true, is a zero-infrastructure privacy tool: no server cooperation, no cryptographic protocols, just background noise in the query stream.

What carries the argument

The carrying mechanism is a client-side automation tool built on a headless browser with human-like typing and randomized delays, which interleaves genuine searches with random queries drawn from language-specific dictionaries. The experiments also rely on Seznam's transparent 'Areas of Interest' feedback, which exposes the day-updated interest profile that other engines keep hidden. The two evaluation metrics are the Jaccard index for set overlap and edit distance for ranked-list reordering, compared between virtual-machine conditions and tested with the Mann-Whitney U statistic.

What would settle it

Re-run the delayed-obfuscation experiment and compare the Delay VM's day-by-day profile trajectory to the Normal VM's using a block bootstrap that preserves autocorrelation; if the Delay VM does not diverge more than the Normal VM's own day-to-day drift after obfuscation begins, the claimed profile overwriting collapses.

Watch

Extended reading notes

Core claim

On Seznam.cz, interleaving genuine queries with randomly generated dictionary queries disrupts the profile the engine builds, while leaving the visible search results largely intact. The paper reports statistically significant differences between a virtual machine whose queries are purely genuine and virtual machines that mix in random Czech, English, French, and other language queries; interest-profile divergence grows with the ratio of decoy to genuine queries and can be produced with a single language (Czech) more effectively than with many. A delayed-obfuscation experiment shows that after five days of ordinary profiling, five more days of injected random queries measurably move the engine's identified interests away from the previously established set. The paper frames this as evidence that profiling pipelines use longer-term frequency and recency signals that are sensitive to query volume, while result ranking reacts to more immediate signals, so the two can be decoupled.

Load-bearing premise

The results assume that the daily interest readings can be treated as independent measurements, even though Seznam's profile is cumulative and each day's reading contains all earlier days' history.

Editorial extensions

If this is right

  • A user can degrade the accuracy of their inferred profile without giving up the usefulness of search results.
  • Profiles are not permanent: already-formed interest sets can be pushed toward different topics by continued decoy traffic.
  • Obfuscation strategy should be tuned to the engine's language domain, since on a Czech-focused engine Czech-only noise beat eight-language noise.
  • Higher decoy-to-genuine ratios, up to the tested 7:1, produce stronger profile disruption, giving users a dial between bandwidth and privacy.
  • The technique requires no changes to search engine infrastructure and can run as a background process during normal browsing.

Reading between the lines

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

  • If the decoupling generalizes, the same technique might be used to actively steer a profile toward chosen interests rather than just scramble it, effectively letting users curate the identity an ad platform sees.
  • Random word queries are likely a lower bound on effectiveness: semantically coherent fake queries that mimic real query structure could defeat entropy-based filters while still diluting the profile.
  • Because Seznam's profile is cumulative, the reported statistical significance may be optimistic; reanalyzing the daily readings with autocorrelation-aware tests would clarify how much of the effect is genuine divergence versus accumulated history.
  • The approach only resists a passive profiler, so the next natural test is whether obfuscation still works when the engine actively filters low-engagement traffic or fingerprints the session.
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

4 major / 6 minor

Summary. The paper proposes a client-side query obfuscation tool that interleaves genuine search queries with randomly generated multilingual queries to disrupt search-engine user profiling. The authors evaluate the tool on Seznam.cz over four experiments, varying language diversity, random-to-genuine query ratio, and the timing of obfuscation onset, while measuring both displayed top-10 search results and Seznam's daily-updated 'Areas of Interest'. The central claim is that displayed search results remain largely stable, while the search engine's identified user interests shift significantly under obfuscation, and that even established profiles can be overwritten. The paper also makes an open-source tool and a real-world measurement dataset available.

Significance. If the central claim is valid, the paper offers practical evidence that lightweight, client-side query obfuscation can meaningfully degrade search-engine profiling without harming search-result utility, and it provides one of the first direct empirical observations of profile-level changes through Seznam.cz's transparent interest panel. The open-source tool and the reproducible experimental workflow are concrete assets. However, the statistical foundation of the claim is currently weak: the experiments use a single VM per condition and the significance tests treat cumulative daily profile readings as independent samples, so the reported p-values likely overstate the strength of the evidence. The descriptive trajectories are suggestive, but the paper does not yet establish that the observed shifts generalize or that the effect is statistically significant in the sense claimed.

major comments (4)
  1. [§5.5, Table 1] The Mann-Whitney U tests in Section 5.5 treat the ten daily 'Areas of Interest' readings per VM as independent samples, but Section 5.4 states that these areas are daily-updated profiles derived from the cumulative search history. The reading on day t therefore contains the history of days 1 through t-1, making consecutive observations strongly autocorrelated within a single VM trajectory. Moreover, each experimental condition is implemented on exactly one VM, so the unit of replication is the VM, not the day. The p-values in Table 1 (e.g., p < 0.0001) are therefore likely artifacts of pseudoreplication and do not support the abstract's claim that identified user interests shift 'significantly' under obfuscation. The analysis should either use a method that accounts for the cumulative/autocorrelated structure (e.g., permutation tests on the full trajectory, time-series models, or block bootstrap) or be explicitly reframed as a descriptive case study without significance claims.
  2. [§5.3, §6] Each experimental condition is instantiated with a single VM and a single Seznam.cz account. There is no replication across accounts, network locations, or time blocks, so the observed differences between conditions cannot be separated from idiosyncratic account-specific or network-specific factors. For example, the Tool VM and Normal VM may differ for reasons unrelated to the obfuscation treatment, such as bot-detection flags, account state, or transient network routing. The manuscript should either add replication or substantially weaken the causal language in Sections 6 and 7; at minimum, provide effect sizes, confidence intervals, and a clear statement that the findings are single-trajectory observations.
  3. [§5.1, §7.6] The experiments were run sequentially over more than 40 days on a university subnet, and the authors themselves note in Section 7.6 that serialized execution may introduce day-to-day search index drift. Because there is no time-matched control condition and no randomization of condition order, temporal drift and other slow-varying confounds are completely aliased with the treatment effects. This undermines cross-experiment comparisons such as 'using only Czech yielded more divergence than using eight languages' and 'a higher ratio led to greater divergence.' The paper should at least present the raw daily series for all VMs side by side and discuss whether the effects are visible within a single time window, or collect interleaved/replicated measurements in future work.
  4. [§6, Table 1] The manuscript performs multiple statistical tests across four experiments and two metrics without any multiple-comparison correction. For instance, the Experiment Four identified-interests p-value of 0.0402 would not survive a simple Bonferroni correction across the ten tests reported in Table 1. In addition, statements such as 'no statistically significant changes in search results' are based on non-significant p-values from single-replicate designs with very low power; absence of evidence is not evidence of absence, and the paper repeatedly interprets these non-significant results as a substantive finding of stability. The paper should either correct for multiple comparisons, or avoid claiming that the null hypothesis of no change is supported for the search-result metrics.
minor comments (6)
  1. [Abstract and Section 1] The abstract and first paragraph contain a spacing error ('Modernsearchengines'), and reference [14] contains the typo '202č' in the year; these should be corrected.
  2. [§4.2] The mode names 'the_tool' and 'queries' are cryptic; renaming them to something like 'obfuscation' and 'targeted' would improve readability.
  3. [§5.2] The description of query timing is ambiguous: the text says '30 queries per day (one every 960 seconds)' and '8-hour periods.' Please clarify whether the 8-hour period is the active window per day and whether the 960-second interval is the median or fixed delay, since this affects reproducibility.
  4. [§6, Figures] The figures are referenced but not fully described in the text; please ensure each figure has a clear caption explaining what is plotted (e.g., median vs. mean, error bars, number of days) and that the Edit Distance scale is explicitly defined.
  5. [§5.5, Table 1] Table 1 reports only p-values; it would be more informative to report medians, interquartile ranges, and effect sizes for the Jaccard Index and Edit Distance under each condition.
  6. [§7.9] The concrete use-case paragraph is useful, but it assumes a 1:3 ratio of genuine to random queries without referencing the experimental results that justify this choice; please tie this scenario explicitly to the measured outcomes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical measurement study whose central claim is evaluated against an external profiling benchmark, not derived from its inputs.

full rationale

This paper is an empirical measurement study rather than a derivation. The central claim, that interleaving random multilingual queries with genuine queries shifts Seznam's displayed 'Areas of Interest', is evaluated by directly comparing VM interest profiles read from Seznam's own transparency interface against a Normal VM baseline. No equation in the paper defines the outcome as a function of the input parameters, no fitted parameter is renamed as a prediction, and no uniqueness theorem or prior work by the same authors is invoked to force the conclusion. The only self-citation, reference [14], is provenance for the tool and does not carry the argument. The fact that Seznam's profile is both the attack surface and the benchmark is an object-level choice, not circularity: the experiment tests whether externally observable profile labels shift under obfuscation. The statistical concern about treating daily cumulative interest readings as independent Mann-Whitney samples is a validity threat, not a circularity, because it does not make the claimed effect equal to its input by construction. Accordingly, no circular step is identified.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its load-bearing assumptions are about the validity of the measurement interface, the passive adversary, the statistical independence of cumulative observations, and cross-engine generalizability. The free parameters are operational settings of the obfuscation tool, none of which are fitted to the outcome; they are chosen by hand and compared directly.

free parameters (6)
  • query length range = 1-3 words
    Random queries were formed from 1-3 dictionary words, a choice inspired by prior work on human query length [6]. The paper does not test sensitivity to this range.
  • random query interval = 320 seconds
    The delay between random queries was chosen to mimic human behavior and avoid carry-over effects; it is a hand-set operational parameter.
  • genuine query interval = 960 seconds
    The interval between profiling queries was set to mimic natural search spacing.
  • random-to-genuine query ratio = 1:1, 1:3, 1:7
    Different ratios were compared to assess dilution strength. The paper concludes 1:7 works best, but this is a direct comparison of chosen values, not a fitted optimum.
  • language pool size = 1, 3, 4, 8 languages
    The number of languages used for random queries was varied across experiments (Czech only vs. eight languages, and the default four: Czech, English, French, Spanish). The finding that Czech-only works better is based on this specific comparison.
  • dictionary wordlists = curated general vocabulary
    The specific words in the language dictionaries were chosen to represent broad topics. No analysis of dictionary composition's effect is provided.
assumptions (4)
  • domain assumption Seznam.cz's 'Areas of Interest' panel accurately reflects the user profile the engine uses for personalization and advertising.
    The entire evaluation depends on this interface being a meaningful proxy for the engine's internal profiling model. The paper asserts this without validation. Introduced in Section 5.4.
  • domain assumption The profiling adversary is passive and does not filter out obvious decoy queries.
    The threat model in Section 3 explicitly assumes the search engine does not actively detect obfuscation. This keeps the evaluation tractable but limits real-world applicability.
  • domain assumption Daily interest profile measurements are independent samples.
    The Mann-Whitney U tests in Section 5.5 compare daily values across VM pairs. Since profiles are cumulative, the readings are autocorrelated, making this assumption false and the p-values unreliable.
  • domain assumption Query-based profiling mechanisms are shared across major search engines.
    The conclusion generalizes the findings beyond Seznam.cz to Google and Bing based on the assumption that the underlying mechanism is similar. This is speculative and untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hiding in Plain Sight: Query Obfuscation via Random Multilingual Searches." pith.science (2026). https://pith.science/paper/NVV5IHVE

@misc{pith2026250604963,
  author       = {Pith},
  title        = {Pith review of: Hiding in Plain Sight: Query Obfuscation via Random Multilingual Searches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NVV5IHVE}},
  note         = {Machine review of arXiv:2506.04963}
}
read the original abstract

Modern search engines extensively personalize results by building detailed user profiles based on query history and behaviour. While personalization can enhance relevance, it introduces privacy risks and can lead to filter bubbles. This paper proposes and evaluates a lightweight, client-side query obfuscation strategy using randomly generated multilingual search queries to disrupt user profiling. Through controlled experiments on the Seznam.cz search engine, we assess the impact of interleaving real queries with obfuscating noise in various language configurations and ratios. Our findings show that while displayed search results remain largely stable, the search engine's identified user interests shift significantly under obfuscation. We further demonstrate that such random queries can prevent accurate profiling and overwrite established user profiles. This study provides practical evidence for query obfuscation as a viable privacy-preserving mechanism and introduces a tool that enables users to autonomously protect their search behaviour without modifying existing infrastructure.

Figures

Figures reproduced from arXiv: 2506.04963 by the authors.

Figure 1
Figure 1. Search Result Similarity (Jaccard Index and Edit Distance) [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Identified Interests Similarity. Note the Edit Distance scale [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Language Experiment: Identified Interests (Jaccard Index - Top, Edit Distance - Bottom). RQ3: Obfuscation Ratio Effects Experiment Three revealed no statistically significant changes in search results, but strong effects on identified interests. A higher random-to-genuine query ra￾tio (1:7) led to greater divergence from the Normal VM compared to a lower ratio (1:1), as illustrated in [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Ratio Experiment: Identified Interests (Jaccard Index - Top, Edit Distance - Bottom) [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Delayed Obfuscation: Changes in Identified Interests [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 19 canonical work pages

  1. [1]

    2023 4th International Conference for Emerging Technology (INCET) pp

    Bhopale, S.D., Sahu, A., Pandyaji, K.: Web services recommendation system us- ing machine learning algorithms. 2023 4th International Conference for Emerging Technology (INCET) pp. 1–7 (2023). https://doi.org/10.1109/INCET57972.2023. 10170205

  2. [2]

    Ethics and Information Technology 15(3), 209–227 (9 2013)

    Bozdag, E.: Bias in algorithmic filtering and personalization. Ethics and Information Technology 15(3), 209–227 (9 2013). https://doi.org/10.1007/ s10676-013-9321-6

  3. [3]

    Online Information Re- view 33(4), 720–744 (2009)

    Domingo-Ferrer, J., Solanas, A., Castellà-Roca, J.: h(k)-private information re- trieval from privacy-uncooperative queryable databases. Online Information Re- view 33(4), 720–744 (2009). https://doi.org/10.1108/14684520910985693

  4. [4]

    IEEE Access7, 144907–144924 (2019)

    Eke, C.I., Norman, A., Shuib, L., Nweke, H.F.: A survey of user profiling: State- of-the-art, challenges, and solutions. IEEE Access7, 144907–144924 (2019). https: //doi.org/10.1109/ACCESS.2019.2944243

  5. [5]

    In: European Conference on Information Retrieval (ECIR)

    Faggioli, G., Ferro, N.: Query obfuscation for information retrieval through differ- ential privacy. In: European Conference on Information Retrieval (ECIR). Lecture Notes in Computer Science, vol. 14608, pp. 278–294 (2024). https://doi.org/10. 1007/978-3-031-56027-9_17

  6. [6]

    517–518 (2005)

    Freund, L., Toms, E.: Understanding the brevity of web queries pp. 517–518 (2005). https://doi.org/10.1002/meet.14504001103

  7. [7]

    In: Proceedings of the 2014 ACM SIGSAC Conference on Com- puter and Communications Security (CCS)

    Gervais, A., Shokri, R., Singla, A., Capkun, S., Lenders, V.: Quantifying web- search privacy. In: Proceedings of the 2014 ACM SIGSAC Conference on Com- puter and Communications Security (CCS). pp. 966–977 (2014). https://doi.org/ 10.1145/2660267.2660367

  8. [8]

    In: Proceedings of the 10th ACM SIGCOMM Conference on Internet Mea- surement

    Guha, S., Cheng, B., Francis, P.: Challenges in measuring online advertising sys- tems. In: Proceedings of the 10th ACM SIGCOMM Conference on Internet Mea- surement. p. 81–87. IMC ’10, Association for Computing Machinery, New York, NY, USA (2010). https://doi.org/10.1145/1879141.1879152, https://doi.org/10. 1145/1879141.1879152 16 Firc et al

Show all 26 references
  1. [9]

    Proceedings of the 22nd internationalconferenceonWorldWideWeb(May2013).https://doi.org/10.1145/ 2488388.2488435

    Hannak, A., Sapieżyński, P., Kakhki, A.M., Krishnamurthy, B., Lazer, D., Mislove, A., Wilson, C.: Measuring personalization of web search. Proceedings of the 22nd internationalconferenceonWorldWideWeb(May2013).https://doi.org/10.1145/ 2488388.2488435

  2. [10]

    IEEE Transactions on Information Forensics and Security18, 2132–2147 (2023)

    Houssiau, F., Liénart, T., Hendrickx, J., de Montjoye, Y.A.: Web privacy: A for- mal adversarial model for query obfuscation. IEEE Transactions on Information Forensics and Security18, 2132–2147 (2023). https://doi.org/10.1109/TIFS.2023. 3262123

  3. [11]

    Lessons from the Identity Trail: Anonymity, Privacy, and Identity in a Networked Society 23 (01 2009)

    Howe, D., NISSENBAUM, H.: Trackmenot: Resisting surveillance in web search. Lessons from the Identity Trail: Anonymity, Privacy, and Identity in a Networked Society 23 (01 2009)

  4. [12]

    International Journal of Intelligent Systems 28(6), 606–622 (2013)

    Juárez, M., Torra, V.: Toward a privacy agent for information retrieval. International Journal of Intelligent Systems 28(6), 606–622 (2013). https: //doi.org/https://doi.org/10.1002/int.21595, https://onlinelibrary.wiley.com/doi/ abs/10.1002/int.21595

  5. [13]

    International Journal of Advance Foundation and Research in Com- puter 1, 2348–4853 (11 2014)

    Kanoje, S., Girase, S., Mukhopadhyay, D.: User profiling trends, techniques and applications. International Journal of Advance Foundation and Research in Com- puter 1, 2348–4853 (11 2014)

  6. [14]

    Mas- ter’s thesis, Brno University of Technology, Brno, Czech republic (202č), https: //www.vut.cz/en/students/final-thesis/detail/153822

    Klusáček, J.: Protection against profiling with random multilingual search. Mas- ter’s thesis, Brno University of Technology, Brno, Czech republic (202č), https: //www.vut.cz/en/students/final-thesis/detail/153822

  7. [15]

    In- ternational Journal of Emerging Trends in Engineering Research8(4), 1410–1416 (Apr 2020)

    Kumar, K.: Privacy protection in personalized web search using obfuscation. In- ternational Journal of Emerging Trends in Engineering Research8(4), 1410–1416 (Apr 2020). https://doi.org/10.30534/ijeter/2020/76842020, http://dx.doi.org/10. 30534/ijeter/2020/76842020

  8. [16]

    Lawrence, S.R.: Personalization of web search results using term, category, and link-based user profiles (9 2012)

  9. [17]

    https://doi.org/10.1515/DMA.2008.036

    Maylybaeva, G.A.: The order of communication complexity of pir-protocols18, 505 – 515 (2008). https://doi.org/10.1515/DMA.2008.036

  10. [18]

    Annals of Cardiac Anaesthesia 22, 67 – 72 (2019)

    Mishra, P., Pandey, C., Singh, U., Gupta, A., Sahu, C., Keshri, A.: Descriptive statistics and normality tests for statistical data. Annals of Cardiac Anaesthesia 22, 67 – 72 (2019). https://doi.org/10.4103/aca.ACA_157_18

  11. [19]

    Korean Journal of Anesthesiology69, 8 – 14 (2016)

    Nahm, F.: Nonparametric statistical tests for the continuous data: the basic con- cept and the practical use. Korean Journal of Anesthesiology69, 8 – 14 (2016). https://doi.org/10.4097/kjae.2016.69.1.8

  12. [20]

    In: Proceedings of the 6th ACM Symposium on Information, Computer and Communications Security

    Peddinti, S.T., Saxena, N.: On the effectiveness of anonymizing networks for web search privacy. In: Proceedings of the 6th ACM Symposium on Information, Computer and Communications Security. p. 483–489. ASIACCS ’11, Association for Computing Machinery, New York, NY, USA (2011...

  13. [21]

    Peddinti,S.T.,Saxena,N.:Websearchqueryprivacy:Evaluatingqueryobfuscation and anonymizing networks. J. Comput. Secur.22(1), 155–199 (jan 2014)

  14. [22]

    International Journal of Computer Applications158(1), 50–54 (2017)

    Punagin, S., Arya, A.: A novel query obfuscation scheme with user controlled pri- vacy and personalization. International Journal of Computer Applications158(1), 50–54 (2017)

  15. [23]

    Inter- national Journal on Document Analysis and Recognition5, 67–85 (2002)

    Schulz, K., Mihov, S.: Fast string correction with levenshtein automata. Inter- national Journal on Document Analysis and Recognition5, 67–85 (2002). https: //doi.org/10.1007/s10032-002-0082-8

  16. [24]

    Social Network Analysis and Mining 10, 1–16 (2020)

    Verma, V., Aggarwal, R.: A comparative analysis of similarity measures akin to the jaccard index in collaborative recommendations: empirical and theoret- Hiding in Plain Sight: Query Obfuscation via Random Multilingual Searches 17 ical perspective. Social Network Analysis and ...

  17. [25]

    IEEE Transactions on Dependable and Secure Computing18(6), 2719–2732 (2021)

    Wei, C., Gu, Q., Ji, S., Chen, W., Wang, Z., Beyah, R.: Ob-wspes: A uniform evaluation system for obfuscation-based web search privacy. IEEE Transactions on Dependable and Secure Computing18(6), 2719–2732 (2021). https://doi.org/10. 1109/TDSC.2019.2962440

  18. [26]

    In: Proceedings of the 16th International Conference on World Wide Web

    Xu, Y., Wang, K., Zhang, B., Chen, Z.: Privacy-enhancing personalized web search. In: Proceedings of the 16th International Conference on World Wide Web. p. 591–600. WWW ’07, Association for Computing Machinery, New York, NY, USA (2007). https://doi.org/10.1145/1242572.1242652...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.