Elite-Attack Corpus: Sybil Review Analysis

• Elite-Attack Corpus is a structured dataset that reconstructs Dianping’s coordinated elite Sybil attacks using raw review metadata, graph structures, campaign windows, and elite user scores.
• It operationalizes detection through a multi-stage pipeline involving collusion graph construction, community labeling, and dynamic campaign window identification.
• The corpus offers actionable insights into temporal collusion, filter bypass effectiveness, and the economic underpinnings of coordinated review manipulation across diverse store categories.

Searching arXiv for the primary paper and closely related elite Sybil attack work. The Elite-Attack Corpus denotes the Dianping-centered review-manipulation corpus implicit in the study of elite Sybil attacks introduced in "Smoke Screener or Straight Shooter: Detecting Elite Sybil Attacks in User-Review Social Networks" (Zheng et al., 2017). It is not a formally named public dataset in the paper, but rather a corpus identity that can be reconstructed from the study’s raw crawl, derived graph structures, community labels, campaign windows, and user-level elite-Sybil scores. The corpus was assembled to characterize and detect a specific class of coordinated reputation attacks in user-review social networks (URSNs): operations in which highly rated accounts that normally appear trustworthy are recruited to post campaign-aligned fake reviews. Within that framing, the corpus captures both the observable review-layer traces of manipulation and the higher-order coordination signals needed to infer communities, campaigns, and elite participants (Zheng et al., 2017).

1. Corpus identity and scope

In the paper’s framing, the corpus consists of four interconnected strata: raw review metadata and store–user linkage; derived collusion networks and community labels; inferred campaign windows; and user-level elite-Sybil scores (Zheng et al., 2017). The study does not assign a formal dataset name, and it was not publicly released at publication time. The authors state, “We will make all of our data used publicly available in the future,” but provide no repository link. Reproduction is therefore possible in principle from the reported statistics, formulas, workflow, and validation procedures, rather than from a published archive (Zheng et al., 2017).

The platform is Dianping, described as by far the most popular URSN service in China in the study. The crawl covers reviews from Jan 1, 2014–Jun 15, 2015, with a reported scale of 10,541,931 reviews, 3,555,154 users, and 32,940 stores; the paper later cites 32,933 stores after crawl completion. Coverage spans 13 categories, including cinemas, hotels, restaurants, entertainment, shopping, and chain stores, with regional granularity via store districts (Zheng et al., 2017).

The corpus is therefore both a behavioral dataset and a detection substrate. It preserves the public review-layer signals needed to build collusion graphs, while also encoding the outputs of a multi-stage inference pipeline. A plausible implication is that the corpus is most useful when treated not as a flat review table but as a layered analytical object in which raw events, graph structure, temporal segmentation, and user scoring remain separable.

2. Elite Sybil attacks as the organizing phenomenon

The corpus is defined by the attack class it was built to expose. Elite Sybil attacks recruit highly rated accounts—illustrated by Yelp “Elite” users or Dianping 4–6-star users—that normally post genuine reviews but occasionally post fake reviews under the direction of a Sybil leader (Zheng et al., 2017). The paper distinguishes two types of elite Sybil accounts: cultivated Sybil accounts that post many non-task reviews as “smoke-screening,” and benign high-rated accounts temporarily “converted” for pay to complete campaign tasks. In both cases, the accounts are orchestrated by a single entity and therefore remain Sybils in the classical sense (Zheng et al., 2017).

The study argues that elite Sybils differ from ordinary Sybils in several ways. They are more active outside campaigns, have a much lower fraction of fake reviews in their history, possess higher user-level star ratings, produce more realistic content often accompanied by photos, and achieve higher platform filter bypass rates. Ordinary Sybils, by contrast, concentrate fake posts in bursts and are filtered more often (Zheng et al., 2017). Relative to organic users, elite Sybils exhibit temporal collusion on target stores and align rating and content with campaign directives; organic users do not show systematic collusion across multiple campaigns (Zheng et al., 2017).

The paper also identifies a hybrid organizational architecture with four actors: customers or overhyped stores, agents or organizers, leaders, and elite Sybil workers. Two leader models are reported. In the leader-supervised model, workers craft content to the leader’s specification and timing; in the leader hands-on model, the leader or customer supplies ready-to-post high-quality text and images (Zheng et al., 2017). The economic structure is explicit: elite accounts can earn up to 20× more than regular Sybils for the same task; cultivating a 3-star account was observed at roughly $6 per account; and tasks can require hundreds of positive reviews for a single store with after-sales guarantees such as re-posting if reviews are filtered (Zheng et al., 2017).

This organizational model is central to the corpus because it explains why the resulting graph is large-scale and sparsely-knit. The corpus is not merely a collection of fake reviews; it is a record of a market-driven coordination system whose signals are diluted by the apparently organic behavior of high-status accounts.

3. Data layers, annotations, and corpus composition

The study operationalizes the corpus through a combination of crawling, community labeling, and elite-user validation. Community labeling was performed on 170 communities by five native Chinese annotators (undergraduates) using majority voting. The outcome was 117 Sybil and 53 benign communities (Zheng et al., 2017). A community was labeled Sybil if at least two of three criteria held: massive filtered reviews by Dianping in the community; duplicate user reviews concentrated on one or two stores with similar content; and spatio-temporal review anomalies consistent with collusion patterns (Zheng et al., 2017). No inter-annotator $\kappa$ is reported. For 74 communities (1,969 users), however, 96.85\% of users’ labels aligned with their community label, which the paper uses to support the binary-community assumption (Zheng et al., 2017).

Elite Sybil user validation occurs after ranking. Human reviewers classify reviews into suspicious versus normal using stricter criteria: presence in many campaign windows, alignment of suspicious reviews with campaign intent such as 5-star boosting, and spatio-temporal separation between suspicious and normal activity (Zheng et al., 2017).

The resulting corpus composition is summarized by the reported counts.

Component	Reported quantity	Notes
Reviews	10,541,931	Dianping crawl, Jan 1, 2014–Jun 15, 2015
Users	3,555,154	Total crawled users
Stores	32,940 / 32,933	Later store count updated after crawl completion
Sybil communities	566	Detected communities
Benign communities	144	Predicted by trained classifier
Regular Sybil users	21,871	User class count
Elite Sybil users	12,292	User class count
Combined Sybil accounts	$\sim 34{,}163$	Versus 3.56M total users
Fake reviews	More than 108,100	Identified as fake

Campaign counts are reported in two forms. The abstract reports 2,164 Sybil campaigns, whereas the temporal analysis section reports 4,162 campaigns after broadly applying the campaign-window algorithm. The latter count is used in duration analyses, including the finding of 466 one-day “ephemeral” campaigns (Zheng et al., 2017). This discrepancy is best treated as a difference between summary reporting and algorithmic expansion rather than as a contradiction in corpus identity.

The paper further reports that 12.37\% of Sybil communities post for chain stores, sometimes across multiple branches of the same brand (Zheng et al., 2017). That detail is significant because it shows that the corpus captures coordinated manipulation not only at the single-store level but also in multi-branch commercial settings.

4. Detection pipeline and formal constructs

The corpus is inseparable from ElsieDet, the three-stage detection system used to derive its key labels and scores (Zheng et al., 2017). The stages are: separating suspicious user groups, identifying campaign windows, and identifying elite Sybil users. Because the corpus includes graph-, campaign-, and user-layer outputs, ElsieDet’s constructs are effectively part of the corpus definition.

Stage 1: separating suspicious user groups

Each review is represented as a tuple $(U, T, S, L)$ for user ID, timestamp, store ID, and star rating. For users $u$ and $v$ with review sets $\mathcal{R}(u)$ and $\mathcal{R}(v)$, the paper defines $P_u(k)=1$ if there exists a review by $v$ such that the two reviews concern the same store, occur within a fixed time window $\Delta T$, and both use extreme ratings in $\sim 34{,}163$0. Otherwise $\sim 34{,}163$1 (Zheng et al., 2017).

Pairwise similarity is then defined as

$\sim 34{,}163$2

An undirected weighted graph $\sim 34{,}163$3 is constructed with an edge $\sim 34{,}163$4 iff $\sim 34{,}163$5, after which the Louvain method partitions the graph into communities (Zheng et al., 2017). Communities are then classified as Sybil or benign using eight features across three families: community-based, similarity-based network, and user-based aggregation features. The eight features are: score deviation, average reviews per store, entropy of chain-store participation, entropy of districts, average pairwise similarity among users in the community, global clustering coefficient, unique reviews ratio, and maximum number of duplicates per user in the community (Zheng et al., 2017). The paper states that no linguistic features were used due to limited effectiveness in URSNs (Zheng et al., 2017).

Stage 2: campaign window identification

For a labeled Sybil community and target store, the input is a weekly series $\sim 34{,}163$6, where $\sim 34{,}163$7 is the number of reviews posted in week $\sim 34{,}163$8 (Zheng et al., 2017). For an interval $\sim 34{,}163$9, let $(U, T, S, L)$0 be the number of weeks with positive review counts and $(U, T, S, L)$1 the number of weeks with zero review counts. The interval is called sparse if $(U, T, S, L)$2 (Zheng et al., 2017).

The algorithm initializes the left and right boundaries and iteratively removes sparse intervals from whichever side contains fewer total reviews, preserving the dense core. When no sparse intervals remain, the residue $(U, T, S, L)$3 is taken as the campaign window (Zheng et al., 2017). The paper’s rationale is that generic burst detection can be confounded by legitimate promotions, whereas this procedure isolates coordinated activity by a labeled Sybil community.

Stage 3: identifying elite Sybil users

For community $(U, T, S, L)$4, let $(U, T, S, L)$5 denote the number of reviews in the $(U, T, S, L)$6th detected window and $(U, T, S, L)$7 the maximum over windows. The window weight is

$(U, T, S, L)$8

For user $(U, T, S, L)$9 in community $u$0,

$u$1

where $u$2 is $u$3’s review count in window $u$4 (Zheng et al., 2017). Participation rate is then sigmoid-normalized within $u$5:

$u$6

Finally, Sybilness across communities is scored by

$u$7

Users are flagged as elite when they do not belong to any clustered community yet exhibit strong campaign participation, for example $u$8 for some $u$9 (Zheng et al., 2017). The paper also defines a per-review suspiciousness score $v$0, which can support downstream throttling or filtering.

These formal definitions turn the corpus into a reproducible analytical framework. The raw layer records only public review metadata, but the corpus becomes operationally meaningful because it preserves the graph edges, community IDs, campaign windows, and user-level scores computed from those definitions.

5. Evaluation and empirical findings

Community classification was evaluated with 5-fold cross-validation on the 170 labeled communities. The best model was SVM (RBF, $v$1, $v$2), achieving Precision 96.74\%, Recall 96.47\%, F1 96.45\%, and AUC 99.42\% (Zheng et al., 2017). Other classifiers—decision tree, GNB, KNN, AdaBoost, random forest—performed slightly worse (Zheng et al., 2017). The predicted composition after classification was 566 Sybil communities (22,324 users) and 144 benign communities (5,222 users) (Zheng et al., 2017).

The paper reports a strong difference in filtered-review distributions: 80\% of Sybil communities have more than 80\% filtered reviews, while benign communities typically remain below 50\% (Zheng et al., 2017). Elite-user validation is also strong by the paper’s manual review protocol: 93.8\% precision on the top-1,000 ranked suspected elite users, and 90.7\% precision on a random 1,000 flagged users (Zheng et al., 2017).

A central empirical finding is filter bypass. The paper states that less than 33.7\% of elite Sybil fake reviews were filtered by Dianping, implying that most bypassed the platform’s existing filter (Zheng et al., 2017). This directly motivates the corpus: elite Sybil behavior is not adequately captured by systems that rely mainly on conspicuous burstiness or obviously low-quality accounts.

The corpus also supports fine-grained temporal analysis. Campaign durations are described as unimodal with a spike at 7 days, but many campaigns are short-lived, including 466 one-day “ephemeral” campaigns (Zheng et al., 2017). An illustrative case is Community 4559, where 33 elite users posted 127 reviews over about two months to a single store, with some workers posting on weekly or monthly schedules (Zheng et al., 2017). The study interprets such scheduling as camouflage of an “organic” cadence while preserving collusion.

Industry-level impact is nontrivial. The study reports campaign involvement in 30.2\% of cinemas, 7.7\% of hotels, and 5.5\% of restaurants, with additional activity in entertainment, wedding, beauty, fitness, and other categories (Zheng et al., 2017). It also notes a hotel that engaged three distinct Sybil communities, with spikes in review counts aligned with elevated star ratings that diminished when fake reviews were removed from the calculation (Zheng et al., 2017). This suggests that the corpus is suitable not only for account-level detection but also for studying market-level manipulation patterns.

6. Early warning, reproducibility, and limitations

The paper argues that the corpus enables early warning when elite Sybil users are monitored over time. Using a 7-day sliding window and the rule at least 7 elite-user reviews on the same store within 7 days, monitoring only elite Sybil users detects approximately 90.40\% of campaigns across the full period (Zheng et al., 2017). Detection using only the initial part of the campaign window achieves 56.77\% in the first quarter, 63.08\% in the first third, and 75.14\% in the first half (Zheng et al., 2017). With an average campaign length of 68 days, the paper concludes that more than half of campaigns can be flagged within the first two weeks (Zheng et al., 2017).

Because no public release accompanied the paper, reproducibility depends on reconstructing the corpus. The reported pathway is explicit: crawl Dianping public pages starting from a small seed set of suspected overhyped stores; expand iteratively through the bipartite store–user graph until reaching a large and diverse store set; store only public review–store metadata $v$3; build the collusion graph via $v$4 and threshold $v$5; cluster with Louvain; label a subset of communities using the stated criteria and majority voting; train the SVM on the eight features; run the campaign-window algorithm per labeled Sybil community; compute $v$6, $v$7, and $v$8; and validate elite users manually (Zheng et al., 2017).

The paper also specifies privacy safeguards. Only public review–store relations were crawled; no usernames, gender, profile photos, or tags were stored; and no fake reviews were posted (Zheng et al., 2017). ElsieDet uses only public metadata and does not rely on IP or device signals, though the paper notes that Dianping could use such private signals internally (Zheng et al., 2017).

Several limitations are explicit. The findings are platform- and region-specific, reflecting Dianping in 2014–2015; annotation is based on rule-of-thumb criteria, with no inter-annotator agreement reported; tactics may drift, requiring re-tuning of $v$9, $\mathcal{R}(u)$0, and the 7-day window/7-review rule; and the feature set excludes linguistic signals by design (Zheng et al., 2017). The paper suggests that adding content signals could help, but only if robust to adversarial text generation. A plausible implication is that the corpus is best treated as a historically grounded benchmark for coordination detection rather than as a timeless representation of review fraud.

7. Significance within URSN security research

The principal contribution of the Elite-Attack Corpus is that it shifts the unit of analysis from isolated suspicious reviews or obviously fraudulent accounts to temporally coordinated, community-weighted participation by seemingly trustworthy users (Zheng et al., 2017). In that respect, it documents a transition from dense, bursty, easily filtered Sybil activity toward sparse, economically organized, reputation-aware manipulation.

The corpus is also notable for the way it couples graph inference with temporal segmentation. Community detection alone is insufficient because elite accounts may escape clustering; campaign detection alone is insufficient because legitimate promotions can also generate bursts. The joint use of collusion graphs, campaign windows, and user-level participation scores yields a corpus whose labels encode both structure and timing (Zheng et al., 2017). This suggests that future URSN corpora addressing coordinated abuse may need to preserve intermediate analytical layers rather than only final account or review labels.

Finally, the corpus provides a concrete operational schema for future dataset design. In the paper’s terms, a reconstructed release would contain a raw layer with all crawled public reviews and store attributes; a graph layer with pairwise similarity, thresholded edges, and community IDs; a label layer with community labels and annotation criteria logs; a campaign layer with weekly series and detected windows; a user layer with $\mathcal{R}(u)$1, $\mathcal{R}(u)$2, $\mathcal{R}(u)$3, $\mathcal{R}(u)$4, and per-review suspiciousness; and an evaluation layer with folds, SVM parameters, and metrics (Zheng et al., 2017). In that sense, the Elite-Attack Corpus is both an empirical record of a Dianping elite-Sybil ecosystem and a template for how coordinated-review-manipulation datasets can be formalized.