A 1946 Psychology Theory Is Now Catching Malicious Websites with 99.89% Accuracy

In the second quarter of 2023 alone, reported phishing and other malicious online incidents topped 1.2 million. By 2025, confirmed data breaches hit a record 12,195, even as the overall incident count fluctuated (Verizon, 2025). And yet the tools meant to stop these attacks are quietly failing at the very moment they're needed most.

The problem isn't that defenders lack clever algorithms. It's that the cleverest algorithms are built on a fragile assumption — that a malicious website looks different from a legitimate one. Machine-learning models learn to spot suspicious URL strings, suspicious word patterns, suspicious character sequences. Attackers, in turn, learn to blur those lines. A URL with a slightly misspelled bank name, a shortened link, a homoglyph that swaps a Latin "o" for a Cyrillic "о" — each cheap obfuscation defeats a model that only knows how to read the surface.

SiNMULI (Gayen et al., 2026) takes a fundamentally different route. Instead of asking what does the malicious URL look like, it asks who links to it, and why. The researchers, from the Indian Institute of Information Technology Guwahati, model the entire web as a signed social network — every website a node, every hyperlink a relationship of trust or distrust — and then borrow a century-old idea from social psychology to guess which relationships are healthy and which are toxic. The result: 99.89% accuracy, 99.62% precision, and 99.80% F1-score, beating the best baseline by 2.4 percentage points while requiring no training data at all and remaining immune to the obfuscation tricks that defeat conventional models.

The core insight is refreshingly old-fashioned. We already know, intuitively, that links are judgments. When a reputable news site links to a source, that's an endorsement. When a spam farm links to a site, it's something else entirely. SiNMULI just makes this intuition mathematically explicit.

The Science

The idea begins with something called balance theory — a concept invented in 1946 by the Austrian-American psychologist Fritz Heider and formalized in 1956 by Dorwin Cartwright and Frank Harary. Heider noticed that humans are deeply uncomfortable with certain configurations of relationships. If your best friend hates someone, you feel social pressure to hate them too. Networks of friends and enemies, he argued, tend to "balance" themselves into stable configurations.

The structure goes like this: take three connected entities — a triad. If all three relationships are trusting, the triad is balanced. If two people distrust each other but both distrust a third, that too is balanced — a "common adversary" arrangement. But a triad where you trust one person and distrust another, while those two trust each other, feels unstable. Cartwright and Harary gave this intuition a crisp mathematical form: a triad is balanced if the product of its three edge signs is positive (Heider, 1946; Cartwright and Harary, 1956).

$$s_{ij}\cdot s_{jk}\cdot s_{ki}>0$$

where each $s_{xy}\in \{+1,-1\}$ is the sign of the relationship between nodes $x$ and $y$ (Gayen et al., 2026). Balanced triads — (+,+,+) mutual trust, or (+,-,-) shared enemy — are stable. The all-negative triad (,,-,-,-) is unbalanced under "strong" balance but tolerated under the "weak" variant that Davis later proposed for hostile cliques.

and

show the two balanced configurations, while

shows the unbalanced one.

Gayen and colleagues reinterpreted the web through this lens. Every website is a node. Every hyperlink from one site to another is a directed edge. The sign of the edge encodes intent: hyperlinks originating from legitimate websites are trust edges $(+1)$, and hyperlinks from malicious websites are distrust edges $(-1)$. "We hypothesize that hyperlink relationships between URLs always have either a positive or a negative aspect, which reveals the intuition behind the back-link generation," the authors write.

The methodology has three stages. First, they crawl the Phishing Websites Dataset (Ariyadasa et al.), extracting every hyperlink from 80,000 retained URLs — 45,000 legitimate (56.25%), 30,000 malicious (37.50%), and 5,000 unlabeled (6.25%). They deliberately discard content, lexical features, and page text; only the structure of who-links-to-whom survives. Second, they build the signed network, removing the 121 exact duplicate URLs and filtering to structurally valid nodes. Third, they deploy a balance-theoretic inference mechanism.

The heart of the method is deceptively simple: a 51% majority rule across incoming links. For any unlabeled domain, they count the incoming positive edges $n_{+}(v)$ and negative edges $n_{-}(v)$. If more than half the incoming links point toward trust, the domain is labeled legitimate; if the incoming distrust edges dominate, it's labeled malicious. Balance theory supplies the machinery to propagate this reasoning across the network, inferring the signs of previously unlabeled edges so that triads remain structurally consistent.

shows the full SiNMULI workflow.

This is the crucial break from convention. Conventional deep-learning detectors are trained on labeled examples and become brittle the moment attackers change their tactics. SiNMULI, by contrast, requires no training data at all. It's pure structural inference — "independence from training data," as the authors put it — which means it cannot "drift" as attacker tactics evolve, because it never learned attacker tactics in the first place.

What They Found

The headline numbers are striking. SiNMULI achieves 99.89% accuracy, 99.62% precision, and a 99.80% F1-score on the benchmark (Gayen et al., 2026). Precision of 99.62% matters enormously in this domain: it means that of every 100 sites flagged as malicious, fewer than half a dozen are false accusations. In a security context, false positives are not a minor nuisance — they erode user trust in the system itself, and they waste the time of the human analysts who must triage every alert.

The accuracy edge over the best-performing baseline is 2.4 percentage points — a meaningful gap at these scales, where a single percent shift can represent hundreds of false negatives across millions of URLs.

But the paper's most compelling evidence is structural, not numeric. The researchers examined how malicious and legitimate websites differ in their network behavior, and the patterns are stark and consistent. Malicious websites show lower centrality scores — they occupy less influential positions in the web's link economy. They have fewer internal links — sparse internal navigation, as if the site were assembled hastily and abandoned. They show sparser clustering coefficients — their neighborhoods are less densely interconnected than legitimate sites, which tend to form rich clusters of mutual referencing. And the distribution of triads across the network confirms the theory: legitimate ecosystems are dominated by balanced configurations that obey the expectations of trust, while malicious neighborhoods are littered with the unbalanced and unstable patterns that balance theory predicts should not persist.

plots the seven distinct triad types on a logarithmic scale.

In other words, malicious websites are not just different words on a page; they are structurally different members of the web's social fabric. They link differently, they cluster differently, and they occupy different positions in the network's hierarchy of trust. This is why the approach holds up where content-based methods collapse: attackers can reshape the text on a page, but they struggle to fake a decade of organic, trustworthy linking behavior.

The paper's class breakdown also hints at the operational reality. The dataset is deliberately imbalanced — 56.25% legitimate, 37.50% malicious, 6.25% unlabeled — which mirrors the messy real world where most domains are benign and a meaningful minority is actively hostile.

Dataset Composition: 80,000 URLs

Label	Value
Legitimate	45,000
Malicious	30,000
Unknown	5,000

Source: Gayen et al. (2026), Phishing Websites Dataset (Ariyadasa et al.). After removing 121 exact duplicates, 80,000 URLs remained: 45,000 legitimate (56.25%), 30,000 malicious (37.50%), 5,000 unknown (6.25%).

below shows this composition.

Why This Changes Things

Every major detection paradigm before SiNMULI has a fundamental vulnerability, and it's worth naming them plainly.

Blacklists — the Google Safe Browsing and PhishTank approach — are fast but perpetually behind. The median phishing URL lives for only a matter of hours, often long before list updates propagate (O'Harrow, 2013; Kumar et al., 2020; Bell and Komisarczuk, 2020). A list can only tell you about threats you've already seen, and zero-hour attacks are, by definition, unseen.

Lexical and heuristic methods — looking at URL length, token patterns, suspicious keywords — are brittle against trivial obfuscation. Shorten the link, insert a token, swap a homoglyph, and the fixed rules break. Joshi et al. (2019) showed ensembles over static lexical features can be accurate, but accuracy on yesterday's attack isn't resilience against tomorrow's.

Machine-learning and deep-learning pipelines — from classic feature-engineered models (Ma et al., 2009; Sahoo et al., 2017) to end-to-end neural architectures like URLNet (Le et al., 2018) and hybrid CNN–GRU hybrids (Gao et al., 2020) — are the current state of the art, and they're powerful. But they're also data- and compute-hungry, suffer from concept drift as attacker tactics evolve, and — critically — are nearly opaque. When a neural network flags a URL as malicious, no one can say exactly why, which is a liability in a domain where organizations need to justify blocking decisions and triage rapidly.

The history here is worth underscoring because it frames why SiNMULI's contribution is not merely incremental. The field has cycled through feature sets — blacklists, lexical, content, machine-learned, deep-learned — and each cycle improved accuracy while inheriting a new brittleness. The graph approaches that came next, including GNNs over heterogeneous graphs (Zhou et al., 2023) and Loopy Belief Propagation on probabilistic graphs (Guo et al., 2025), recognized that relationships matter but introduced their own costs: large-scale graph construction, streaming updates, noisy edges.

What SiNMULI adds is a social-theoretic prior — a principled rule about how trust structures the world, rather than an empirical pattern learned from data. Balance theory is not a model trained on phishing examples; it's a statement about the structural grammar of trust itself. That's why the approach is simultaneously lightweight, interpretable, and adversarial-robust. There's no black box to reverse-engineer, no feature space to obfuscate, no training distribution to drift away from.

It also upends a common assumption in security: that better detection requires more sophisticated models. Here, the opposite is true. The most sophisticated ML models in the paper are the ones being beaten by a handful of balance-theoretic rules operating on nothing but link structure.

There are genuine limits, and the authors are candid about them. Crawling malicious URLs is hard: many live only briefly, and evasive techniques — bot detection, CAPTCHAs, HTTP 403/404 errors, login-gated content, JavaScript-rendered links — thwart automated extraction. The researchers had to discard thousands of unparsable or structurally insufficient URLs before filtering down to the 80,000-node network. This means the approach depends on a mature, well-linked web of evidence; a brand-new malicious domain with almost no backlinks is hard to classify by structure alone, because structure is precisely what it lacks. The 6.25% "unknown" class in the dataset reflects this reality.

Class Shares in the SiNMULI Dataset

Label	Value
Legitimate	56.25
Malicious	37.5
Unknown	6.25

Source: Gayen et al. (2026). The unknown class represents domains with insufficient or ambiguous backlink structure that the balance-theoretic inference mechanism must resolve.

below illustrates the fraction of nodes that remain genuinely ambiguous.

There's also a conceptual question the paper raises without fully resolving: who vouches for the vouchers? The method seeds trust from a set of known-legitimate and known-malicious nodes, then propagates. If the seed labels are wrong — if a "legitimate" node is actually a sophisticated long-lived scam — the error propagates through the network's trust assumptions. Balance theory is robust against noise in structure, but it inherits the quality of its priors. This is a standard challenge for all trust-inference systems, and it's worth watching in future work.

What's Next

The most exciting implication is that SiNMULI opens the door to a family of network-centric defenses that do not compete on model size — that don't need to train, that don't need labeled datasets, and that don't suffer from the arms race of obfuscation. If structure-based inference works for malicious URLs, the same signed-network framing could plausibly extend to other adversarial ecosystems: fake social-media accounts, disinformation networks, coordinated inauthentic behavior, spam infrastructure, even fraud rings that share the same IP or hosting ecosystem.

The authors flag several near-term directions. One is combining their structural signal with lightweight content features to handle the sparse-neighborhood cold-start problem — hybridizing the network's robustness with the traditional methods' ability to say something about isolated nodes. Another is scaling the approach to streaming settings, where the web graph evolves continuously and new domains must be classified in near-real time rather than in offline batches. A third is exploring status theory — the alternative social-theoretic framework introduced by Leskovec, Huttenlocher, and Kleinberg (2010), which treats relationships as hierarchical judgments of status rather than binary trust — to capture the directed, asymmetric flavor of hyperlink relationships more completely. The paper develops both theories but leans on balance for classification; status theory may refine the sign-propagation mechanism further.

There's also a deeper research question lurking beneath the results. The fact that balance theory — an idea conceived to explain why your friendships feel uncomfortable — works so well at distinguishing malicious from legitimate websites is itself evidence that the web is not a random tangle of links but a social structure with grammar. Like the human networks that inspired it, the web rewards consistent, trustworthy clustering and punishes the unstable configurations that malicious actors create. That's a profound statement about the nature of the medium, and it suggests that social-network theory — long a tool for understanding people — has an untapped role in understanding machines.

For now, SiNMULI stands as a reminder that the best defense is sometimes the oldest insight. We spent decades teaching machines to read the surface of the web. The next generation of cybersecurity may instead teach them to read its relationships — the invisible web of trust and distrust that links every site to every other, and that even the most careful attacker struggles to fake.