{"id":"5e23bc1b-d0d1-4f6d-b727-ccebb4e52c82","arxiv_id":"2504.16836","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A mirror-aware crawl of more than 24,000 Tor hidden-service landing pages finds roughly 82% are replicated content, and ignoring mirrors distorts dark web topology and category estimates.","lead":"This paper measures the Tor hidden-service network with a custom crawler, finding that about 82% of visited pages are copies of other pages. It also shows that ignoring such mirrors skews previous estimates of what the dark web actually contains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 82% mirror share and 74.4% cybercrime share are computed from a deviant-keyword-seeded, landing-page-only crawl; without a neutral-seed comparison or statistical correction they cannot be attributed to the Tor network as a whole.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the crawled subset is not demonstrated to represent the Tor network as a whole. The mirror-detection precision validation and the manually confirmed CSAM findings are real independent support, but they establish that the detector is accurate on the pages it examines, not that the sample is representative. A mirror rate of 82% in a deviant-keyword-seeded crawl can be correct within the sample and still not support the abstract's network-level claim if mirror-heavy counterfeit sites are over-represented by construction. The proposed neutral-seed experiment would settle this directly. Because this is a condition on the interpretation of the headline numbers rather than a demonstration that the construction is wrong, I would keep the reader's CONDITIONAL verdict and add the neutral-seed re-crawl to the list of conditions (alongside artifact release, numeric consistency, and confidence intervals). No change to the verdict is needed.","tokens_in":21566,"tokens_out":9269,"duration_ms":86720,"concrete_test":"Re-run the identical Mimir pipeline (same landing-page-only crawling policy, mirror-detection thresholds, and classifier) starting from at least two independent seed sets that do not use underground-forum keywords: (a) a broad public .onion index such as Ahmia's recently indexed list or the DUTA-10K corpus, and (b) a known set of benign/privacy/activist .onion services. Compare the aggregate mirror rate, the category ranking, and the cybercrime share across the three seed sets. If the mirror rate remains ≈82% and Counterfeit still dominates after mirror filtering, the representativeness concern is refuted; if the rate drops materially or the category ranking changes, the headline numbers must be restated as conditional on the deviant seeding strategy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline percentages are not robust to the crawl's selection bias. Seeds are generated in §II-B by querying Tor search engines with TF-IDF keywords extracted from underground surface forums, plus a small OSINT set, and the crawler follows only external .onion links from landing pages. This deliberately targets deviant content, so the resulting 24,911-site snapshot (§III-A) is a deviant-biased sample, not a representative draw from Tor. The bias directly inflates the aggregate mirror rate: §III-D reports 20,790/24,911 sites as mirrors, but this is driven by heavily mirrored counterfeit operations—57 unique sites have more than 100 mirrors, two have over 1,100, and Table VIII assigns 13,972 of the 20,790 mirrors to Counterfeit. If the seed keywords over-sample the counterfeit mirror ecosystem, both the 82-83% mirror rate and the 74.4% cybercrime share are mechanically inflated. The paper acknowledges the coverage limitation in §V-A, but the abstract and §III generalize the numbers to the Dark Web. No confidence intervals or independent-seed validation are provided, so the reader cannot distinguish a property of the Tor network from a property of the deviant seeding strategy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Mimir, a crawler for Tor hidden services, and uses it to collect landing pages from 24,911 accessible .onion sites over an eight-month period. The authors propose a mirror-detection heuristic combining MD5 exact matching with CTPH fuzzy hashing, using a weighted score over HTML content and structure, and report that roughly 82--83% of crawled landing pages are mirrors. They also train a logistic-regression content classifier on the Duta-10K dataset and report that 74.4% of pages relate to cybercrime (Counterfeit, Drugs, Hacking, Locked). A case study identifies 159--165 sites distributing child abuse material, with a subset still online at the time of writing. The central claim is that prior large-scale Tor measurements, which ignore mirrors, give a biased picture of Dark Web topology and content prevalence.","tokens_in":21856,"tokens_out":5107,"duration_ms":45907,"significance":"If the headline numbers were robust, the paper would make a valuable contribution: it would be one of the first mirror-aware longitudinal measurements of Tor, with a systematic seeding methodology and a plausible argument that mirror duplication distorts category-prevalence estimates. The strengths include the scale of the crawl, the explicit attempt to validate mirror detection on 408 manually checked sites, the benchmark against MinHash and SimHash, and the ethically careful handling of illegal content with reporting to law enforcement. However, the significance is conditional because the main quantitative claims rest on a deviant-focused, shallow crawl whose representativeness is not established, and because several numbers in the paper are internally inconsistent. The methodology is a useful advance, but the paper currently overstates the generality of its 82% mirror and 74.4% cybercrime figures.","major_comments":[{"comment":"The headline mirror rate and cybercrime share are computed over a sample that is explicitly biased toward deviant content, so they cannot be generalized to the Tor network as a whole. Seeds are obtained by querying Tor search engines with TF-IDF keywords extracted from underground surface forums, plus a small OSINT set, and the crawler visits only landing pages and follows external .onion links. Section V-A acknowledges a coverage limitation, but the abstract and §III-D/E present the 82--83% mirror rate and 74.4% cybercrime share as properties of the Dark Web. A neutral-seed comparison, a statistical correction for the sampling design, or a consistent reframing of the claims as sample-specific is needed before these percentages can support the paper's conclusions about prior work being biased.","section":"§II-B, §III-A, §V-A"},{"comment":"Several central quantitative statements are internally inconsistent. Section III-D reports 24,911 sites, 20,790 mirrors, 4,008 unique sites and 113 surface links, but 20,790/24,911 is 83.5%, whereas the abstract and introduction say '≈82%' and the conclusion says 83%. Table III is not clearly formatted and does not reconcile the seed, site, mirror, and unreachable counts. Table VIII lists 1,571 unique cybercrime sites and 16,840 mirror cybercrime sites, which sum to 18,411, not the reported 18,474, and the corresponding percentages do not match the stated totals. These inconsistencies directly affect the paper's central quantitative claims and must be corrected.","section":"§III-D, Table III, Table VIII"},{"comment":"The mirror-detection validation is incomplete and partly circular. The weights Wc=0.7 and Ws=0.3 and the 0.9 similarity threshold are described as empirically tuned to detect mirrors, and the benchmark in Table VI is built from 1,000 'real unique-mirror pairs' and 1,000 non-mirror pairs, but the text does not specify how the ground truth for those pairs was obtained independently of the algorithm. Manual validation in §III-D covers only pages with at least one mirror (408 sites), not non-mirror pages, so false negatives are not measured. As a result, the reported 97--99% precision does not establish that the 0.9 threshold and the chosen weights correctly separate mirrors from non-mirrors across the whole dataset. An independent ground-truth set with both positive and negative pairs is needed.","section":"§II-C, §III-D, Table VI"},{"comment":"The 74.4% cybercrime figure is not well supported. The classifier is trained on only 11 categories with up to 200 samples each from Duta-10K, applied only to English sites, and includes 'Locked' (sites closed by authorities) as a cybercrime category, which conflates takedowns with criminal content. The text itself notes the figures are a lower bound, but the abstract and §V-B state the 74.4% as a factual property of the dataset. The paper should report confidence intervals for the classification-based prevalence estimates, state the denominator explicitly (unique vs. all sites, English vs. all languages), and discuss how classifier error propagates to the aggregate percentage.","section":"§II-E, §III-E, Table VIII"},{"comment":"The child-abuse case-study numbers are inconsistent. Section IV reports 180 pages found, 165 confirmed unique sites (91.66% accuracy), and then says examining mirrors increased detection to 505 pages, described as a '306% increase'. Section V-B reports 159 unique sites and 505 mirrors. The ratios do not work: 505/180 is about 2.8, not a 306% increase, and 159 differs from 165 without explanation. Since the case study is presented as a concrete contribution and is cited in the abstract and conclusions, these figures need to be reconciled.","section":"§IV, §V-B"}],"minor_comments":[{"comment":"The table title reads 'Duplication Algorithms Benchmark' but the table compares machine-learning classifiers; it should be titled 'Classification Benchmark' or similar.","section":"Table VII"},{"comment":"The table layout is difficult to parse; the row 'Base 4,121 6,816' appears to mix seeds and sites, and the total 39,095 is not derived from the surrounding numbers in an obvious way.","section":"Table III"},{"comment":"The sentence 'We manually analyze English pages with at least one mirror (450 and 53.38% in total)' is unclear: 53.38% of what total? The subsequent removal of 43 sites from 450 would leave 407, but the text says 408 sites were validated; please clarify.","section":"§III-D"},{"comment":"The text states the algorithm detects '801 mirrors (355 identified with CTPH and 424 with MD5)', but 355 + 424 = 779; one of these numbers is incorrect.","section":"§III-D"},{"comment":"The sentence about hacking pages being inflated 'by ≈2,966%' is not supported by the data presented in Table VIII or Figure 6; the basis for this percentage should be explained or removed.","section":"§V-B"},{"comment":"The decimal separator is used inconsistently ('91,66%' in §IV, '91.66%' in §V-B); please standardize to the journal's convention.","section":"§IV"},{"comment":"The row for [6] reads '[6] [1]', which appears to be a formatting artifact; the row should be cleaned up.","section":"Table XI"}],"recommendation":"major_revision","confidential_remarks":"This is an extended version of a paper already published in IEEE TIFS. The current manuscript does not clearly delineate which results are new relative to the shorter version, which is relevant for an extended-version submission. The main technical core (the crawler and the mirror-detection idea) is sound and publishable, but the aggregate claims need substantial rework: either add a neutral-seed control experiment or carefully restrict the conclusions to the deviant-focused sample. I would not recommend rejection, because the issues are addressable within the scope of a revision, but the internal numeric inconsistencies must be fixed before the paper can be considered reliable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it builds a real system and ships a genuinely new measurement angle: it detects mirror .onion sites at 25k scale and shows that ignoring them badly distorts content-prevalence rankings. Second, the headline number—82% of the Dark Web is mirrors—is a property of their crawler's deviant-biased seed set, not necessarily of Tor as a whole. The paper acknowledges coverage limits in V-A but the abstract doesn't.\n\nWhat's good: the seed extraction from underground forum keywords via TF-IDF is a nice idea and, per their ablation, beats manual seeds in coverage. The mirror detection design (MD5 + CTPH, weighted content/scheme similarity) is sensible, and the manual validation of 408 sites at 97% precision is real evidence. The benchmark against MinHash/SimHash also shows they thought about baselines. The CSAM case study, while small, is responsibly handled and the law enforcement reporting is a plus.\n\nWhere it gets soft. The selection bias is load-bearing. Seeds are derived from deviant keywords, so any aggregate percentage—mirror share, category distribution, 74.4% cybercrime—is conditional on that sampling. No neutral-seed control is run, no confidence intervals are given. The paper says 'our dataset' in places but the abstract and occasional 'Dark Web' phrasing generalize. That overreach is fixable in revision. The numbers are also internally sloppy: abstract says 82%, conclusion says 83%, Section V-B says around 82%; 159 vs 165 vs 180 for the CSAM sites; the benchmark table has FP counts that dwarf the sample size. These inconsistencies are annoying, not fatal. The mirror-detection weights and threshold were tuned on data that may include the test pairs; the manual validation mitigates but doesn't fully address this. No code or data release is mentioned, which for a measurement paper is a real weakness.\n\nWho is this for? Dark web measurement researchers and anyone doing Tor content analysis. It deserves a serious referee—this isn't desk-reject material. But if I were the editor, I'd ask for major revision: release the crawler and validation data, add a neutral-seed comparison or at least frame the results as 'under deviant seeding', fix the numeric inconsistencies, and add CIs to the headline percentages. The core idea is sound; the presentation currently overclaims.","headline":"A mirror-aware Tor measurement with a real contribution and a selection-bias problem in the headline percentages; the core idea is sound but the abstract overclaims.","tokens_in":22358,"tokens_out":3487,"would_cite":false,"duration_ms":30923,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An eight-month crawl of Tor hidden services finds that about 82% of landing pages are mirrors—exact or near-exact copies—so ignoring duplicates misreads the dark web's size, shape, and content mix.","keywords":["Tor network measurement","hidden services","mirror detection","dark web content analysis","onion services","cybercrime measurement","web crawling"],"falsifier":"Run the same Mimir crawler and mirror detector starting from seeds built from non-deviant topics (for example news, email, or privacy) or from an independent random sample of live onion addresses; if the mirror share falls well below 82%, the headline figure is an artifact of deviant-focused seeding rather than a property of Tor as a whole.","tokens_in":21366,"feed_emoji":"🪞","tokens_out":7516,"duration_ms":66108,"temperature":0.7,"pith_summary":"This paper reports an eight-month crawl of Tor hidden services, carried out by a crawler named Mimir, and argues that about 82% of the landing pages it reached are mirrors—exact or near-exact copies of other pages. It claims that earlier large-scale Tor measurements that ignored these duplicates misread both the topology of the dark web and the distribution of content. To support that, the paper combines a systematic seeding method (keywords derived from underground-forum discussions, fed into Tor search engines), a hybrid hash algorithm for detecting mirrors, and a logistic-regression classifier for content categories. The mirror-aware view changes the picture: counterfeit and pornographic sites dominate among unique pages, and categories like hacking fall in the rankings once copies are removed. A case study on child-abuse material illustrates the practical payoff, detecting 159 unique sites and 505 pages including mirrors, which were reported to law enforcement.","feed_headline":"Most of the dark web is mirrors, eight-month crawl finds","feed_subtitle":"82% of landing pages were copies, which flips earlier rankings of hacking, porn, and counterfeit sites.","key_machinery":"The load-bearing mechanism is a two-stage mirror-detection pipeline. MD5 catches exact duplicates; for non-identical pages, a fuzzy hash (CTPH) compares full HTML, and when languages differ only the HTML tag structure is compared, with a weighted similarity score (0.7 on content, 0.3 on structure) and a 0.9 threshold for declaring a mirror. Around this sits a shallow breadth-first crawler that follows only external .onion links from landing pages, an automatic seeding stage that derives TF-IDF keywords from underground surface-forum titles and queries Tor search engines, and a logistic-regression content classifier trained on balanced categories from the labeled Duta-10K dataset. The paper validates the classifier on a held-out set (86% accuracy) and benchmarks the mirror detector against MinHash and SimHash, reporting higher precision (0.99 versus 0.40–0.61) because those methods repeatedly match one page to many representatives.","core_discovery":"The central discovery is that replicated content is the norm rather than the exception in the Tor hidden-service ecosystem as reached by this crawl. Of 24,911 landing pages collected, 20,790 (about 83%) were detected as mirrors, leaving 4,008 unique onion sites; slightly more than half of the mirrors are exact MD5 matches, and the rest differ only in small ways such as language, prices, or cryptocurrency wallet addresses. From this the paper concludes that any content- or topology-driven measurement of the dark web that does not filter mirrors gives a distorted view. Concretely, without mirror filtering, hacking appears as a top category; after filtering, counterfeit is the largest category and porn and crypto move up. The classifier puts 38% of unique sites in explicitly cybercrime-related categories (Counterfeit, Drugs, Hacking, Locked), but 74.4% of all pages including mirrors fall into those categories, indicating that cybercrime operators mirror their services far more than other operators do.","pith_inferences":["The mirror-aware view suggests a standard practice for future dark-web measurement: report both raw and unique-site statistics, since the gap between them is itself a signal of how aggressively content is replicated.","The same seeding-plus-mirror-detection recipe could be applied to other anonymous or decentralized networks without central indexes, where duplicate content may inflate apparent ecosystem size.","A testable extension is a mirror-churn study: re-crawling the same unique sites over time to measure how quickly replicas appear and disappear would turn this static snapshot into a dynamic model of mirroring behavior.","If the mirror rate stayed near 82% under neutral seeding, mirroring would be a structural property of Tor; if it dropped sharply, prior category-prevalence results would only be wrong for deviant-focused crawls, not for Tor in general."],"forward_implications":["Prior large-scale Tor measurements that ignore mirrors misstate both topology and content distribution; mirror filtering changes category rankings, e.g., hacking is not second-most prevalent once duplicates are removed.","Mirror-aware analysis lets investigators focus deeper inspection on unique sites: only 5.8% of counterfeit-classified sites and 10% of hacking-classified sites need in-depth review.","Because cybercrime operators mirror heavily (81.32% of detected mirrors fall in cybercrime-related categories), mirror detection can help track services that survive takedowns through replicas, as in the child-abuse case where 159 unique sites were detected and some remained operative years later.","Systematic keyword seeding from underground forums covers all sites reachable from manual seeds and adds about 81.85% more coverage, suggesting that the method scales beyond hand-crafted seed lists."],"supporting_citations":[{"why":"Supplies the labeled Duta-10K corpus whose 11 categories ground the content classifier.","marker":"[2]"},{"why":"Provides the earlier large-scale Tor crawl used as the coverage baseline and as an example of mirror-blind measurement.","marker":"[6]"},{"why":"Earlier topology study of Tor hidden services that the paper compares against for subgraph structure.","marker":"[7]"},{"why":"Prior characterization of redundancy in .onion services that motivates the mirror-detection analysis.","marker":"[10]"},{"why":"Defines the CTPH fuzzy-hashing algorithm that the mirror detector applies to HTML, schemes, and content.","marker":"[23]"},{"why":"DarkBERT gives both a classification baseline and a MinHash-based mirror-detection rate that the paper benchmarks against.","marker":"[20]"},{"why":"Defines MinHash, one of the two duplicate-detection baselines in the mirror-detection benchmark.","marker":"[9]"}],"fun_headline_variants":["Dark web mostly mirrors, crawl of 25k sites finds","Tor study: 83% of dark web pages are copies","Dark web replication: 4 in 5 pages are mirrors","Mirrors dominate dark web, new crawl reveals","Dark web is mostly mirrors, 25k-site crawl shows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The crawled subset is representative of the Tor network as a whole: seeds come from underground-forum keywords queried in Tor search engines, and only landing pages are crawled, so the mirror rate and cybercrime share could be inflated if heavily mirrored sites such as counterfeit shops are overrepresented in that sampling path.","fun_headline_variants_meta":{"raw":{"variants":["Dark web mostly mirrors, crawl of 25k sites finds","Tor study: 83% of dark web pages are copies","Dark web replication: 4 in 5 pages are mirrors","Mirrors dominate dark web, new crawl reveals","Dark web is mostly mirrors, 25k-site crawl shows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1556,"prompt_tokens":1056,"completion_tokens":500,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":417}},"tokens_in":672,"tokens_out":500,"duration_ms":4848,"temperature":1.0,"reasoning_tokens":417,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:54:24.629585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Mimir crawler and mirror detector starting from seeds built from non-deviant topics (for example news, email, or privacy) or from an independent random sample of live onion addresses; if the mirror share falls well below 82%, the headline figure is an artifact of deviant-focused seeding rather than a property of Tor as a whole.","supporting_citations":[{"cited_title":"Torank: Identifying the most influential suspicious domains in the tor network,","cited_arxiv_id":null,"evidence_quote":"Supplies the labeled Duta-10K corpus whose 11 categories ground the content classifier."},{"cited_title":"Onion under microscope: An in-depth analysis of the tor web,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier large-scale Tor crawl used as the coverage baseline and as an example of mirror-blind measurement."},{"cited_title":"Identifying almost identical files using context triggered piecewise hashing,","cited_arxiv_id":null,"evidence_quote":"Defines the CTPH fuzzy-hashing algorithm that the mirror detector applies to HTML, schemes, and content."}],"review_version":1}