TraceTarnish: Stylometry & Alloy Tarnish

• TraceTarnish is a dual-framework addressing both adversarial stylometry for text anonymization and descriptor-based prediction of early-stage alloy tarnishing.
• The stylometry pipeline employs round-trip translation, paraphrasing, and Unicode injection to disrupt forensic signatures while retaining natural language semantics.
• The alloy tarnish framework leverages binding energy descriptors and d-band center analysis to forecast chemisorption-driven corrosion on transition-metal surfaces.

TraceTarnish refers to two rigorously defined frameworks in contemporary research: an adversarial stylometry pipeline for textual authorship anonymization and a descriptor-based workflow for predicting the propensity of transition-metal alloys to undergo early-stage tarnishing or corrosion via oxygen and sulfur binding energetics. Both paradigms are united by their focus on the systematic alteration, monitoring, or interpretation of “traces”—stylometric fingerprints in text or physicochemical signatures on metal surfaces—and the means to either erase or reveal these traces through adversarial or diagnostic methodologies (Dilworth, 3 Dec 2025, Tiwari et al., 2021).

1. Adversarial Stylometry and the TraceTarnish Pipeline

TraceTarnish in the domain of computational linguistics designates an attack script developed to anonymize text-based messages by leveraging adversarial stylometry. The primary objective is, given a text $T$ authored by a target, to output an anonymized version $T'$ that (a) preserves the semantic content, (b) corrupts stylometric fingerprints specific to the author, and (c) embeds subtle Unicode steganography such that downstream NLP pipelines may misinterpret or misparse the text. The threat model assumes a defender employing stylometric classification (e.g., a Random Forest classifier over high-dimensional “StyloMetrix” features) to attribute text, and an adversary with access to the original text, paraphrasing, and Unicode injection tools (Dilworth, 3 Dec 2025).

The foundational TraceTarnish algorithm comprises three phases:

1. Round-trip translation via machine translation (e.g., English → French → English), introducing paraphrastic drift.
2. Paraphrasing using rule-based or neural paraphrasers for further obfuscation.
3. Steganographic injection of zero-width Unicode characters (e.g., U+200B) within tokens to introduce imperceptible noise.

This multi-stage transformation results in messages with disrupted stylometric distributions while remaining legible to human readers or most standard NLP workflows.

2. Data Acquisition, Processing, and Stylometric Feature Extraction

TraceTarnish’s empirical validation utilized the Kaggle Reddit comments dataset. To ensure clear attribution boundaries, 100 comments were uniformly sub-sampled such that each came from a unique Reddit user, thus precluding author-mixing. Preprocessing steps included Unicode normalization (NFC), lowercasing, URL and user-mention removal, emoji/markup standardization, whitespace normalization, length control to a 100-token window, and tokenization via standard libraries (e.g., NLTK’s WordPunctTokenizer). For each redact-anonymize cycle, paired records were generated: label 0 for raw (“NANON”), label 1 for anonymized (“ANON”) (Dilworth, 3 Dec 2025).

Feature extraction leveraged StyloMetrix v1.2, which computes 196 stylometric metrics distributed over five categories: basic counts, lexical features, syntactic features, morphological features, and statistical/lexical richness indices. Key metrics and definitions relevant to TraceTarnish include:

• Function-word frequency: $$L_\mathrm{FUNC\_A} = \frac{\#\,\text{function-word tokens}}{N}$$
• Function-word types ratio: $L_\mathrm{FUNC\_T} = \frac{|V_\mathrm{func}|}{|V|}$
• Content-word frequency: $$L_\mathrm{CONT\_A} = \frac{\#\,\text{content-word tokens}}{N}$$
• Content-word types ratio: $L_\mathrm{CONT\_T} = \frac{|V_\mathrm{cont}|}{|V|}$
• Type-Token Ratio (lemmas): $$ST_\mathrm{TYPE\_TOKEN\_RATIO\_LEMMAS} = \frac{|V|}{N}$$

where $N$ is the total token count, $V$ is the set of distinct word lemmas, and $V_\mathrm{func}$, $V_\mathrm{cont}$ are the sets of function and content word lemmas, respectively.

3. Information Gain Selection and Forensic Beacons

Each stylometric feature was evaluated for discriminative power against the binary anonymization label using the Information Gain criterion:

$IG(X;Y) = H(Y) - H(Y|X)$

where $H(Y) = -\sum_{y\in\{0,1\}} p(y) \log_2 p(y)$ and $H(Y|X) = -\sum_x p(x) \sum_y p(y|x) \log_2 p(y|x)$.

Five features exhibited Information Gain $>0.5$ and thus formed the critical “forensic beacon” for attack/defense detection:

Feature	IG Value
ST_TYPE_TOKEN_RATIO_LEMMAS	0.6748
L_CONT_T	0.5987
L_FUNC_A	0.5624
L_CONT_A	0.5421
L_FUNC_T	0.5332

These stylometric cues enable not only attribution by defenders but also detection of adversarial obfuscation, provided both pre- and post-transformation texts are available for comparison.

4. Evaluation, Limitations, and Counter-Forensics

Initial empirical evaluation using a Random Forest classifier (100 trees, default scikit-learn settings) on balanced raw/anonymized datasets achieved perfect separation: 100% accuracy via 5-fold cross-validation. Each anonymized TraceTarnish output is thus perfectly detectable in StyloMetrix feature space, meaning that the anonymization pipeline itself becomes a distinctive indicator of compromise unless counter-measures are taken (Dilworth, 3 Dec 2025).

To address this, a “tuned” (v2) pipeline incorporates adversarial imitation via an offline LLM (“Self-hosted Grammarly”), which rewrites the post-anonymization output to restore function/content word distributions and type-token ratios to approximate corpus-level means. Constraints leverage direct manipulation: e.g., insertion of low-frequency function words, synonym chaining for type balance, and TTR adjustment via lexical repetition or token merging. This enhancement aims to dampen forensic beacons, with an anticipated drop in attack detectability from 100% to 50–70%. Quantitative outcomes for this modification remain subjects of ongoing research.

5. TraceTarnish in Materials Science: Alloy Tarnish Prediction

In parallel, TraceTarnish serves as a conceptual and quantitative workflow for alloy susceptibility to early-stage tarnishing. Here, the principal substrates are transition-metal alloys and the threat is initial O or S chemisorption, modeled using the Newns-Anderson impurity model. The key descriptor is the binding energy $E_b$ (in eV) for O or S on metal surfaces, which captures the thermodynamic driving force for oxide/sulfide nucleation (Tiwari et al., 2021).

Theoretical formalism involves:

• Original NA model Hamiltonian:

$$H_{NA} = \varepsilon_d\, d^\dagger d + \varepsilon_a\, a^\dagger a + V_{ad}(a^\dagger d + d^\dagger a)$$

• Binding energy:

$$E_b = -(1-f)[W_{ad} - \Delta_{ad}], \quad W_{ad} = \sqrt{\Delta_{ad}^2 + 4V_{ad}^2}$$

where $f$ is the fractional d-band filling, $\varepsilon_d$ the d-band center, $\varepsilon_a$ the adsorbate level, and $V_{ad}$ the hybridization parameter. A constant $E_{sp}$ corrects for s–p band contributions.

Fit to DFT data for O and S on 4d metals yields $\sim0.5$ eV (O) and $\sim0.7$ eV (S) accuracy using only three fit parameters. Binding energies, extracted via this descriptor, span –9 to –4 eV (O) and –7 to –3 eV (S) for single-phase metals; binary alloys show a broader range depending on composition and surface site.

6. Design Rules and Diagnostic Monitoring of Tarnish

Alloy design informed by the TraceTarnish descriptor targets late-transition-metal-like $d$-band centers to achieve $E_b(\mathrm{O}) > -4.0$ eV and $E_b(\mathrm{S}) > -3.5$ eV, greatly reducing the propensity for initial chemisorption and stoppage of the corrosion cascade. Pure early-transition-metal sites (Ti, Zr, Y, Hf, Nb) are avoided unless electronically diluted by over 80% late-TM neighbors. Selected low-reactivity compositions include: Cu$_2$NiZn, Ni$_5$Sc, Ni$_3$Zr, Ni$_4$W, Cu$_5$Y, Pd$_3$Ta, Pt$_3$Ta, AuCuZn$_2$, AgAuZn$_2$, Ag$_2$AuCd. For tertiary systems, the least-reactive include Cu$_2$NiZn and AuCuZn$_2$.

Tarnish monitoring exploits spectroscopic shifts in the $d$-band center ($\varepsilon_d$) as proxies for $E_b$, measurable via UPS/XPS or ambient pressure XPS and plasmon resonance. The diagnostic “TraceTarnish index” (Editor's term) aggregates these observables to flag early-stage corrosion or facilitate alloy library screening (Tiwari et al., 2021).

7. Synthesis and Cross-Domain Implications

Both paradigms of TraceTarnish—stylometric adversarial transformation and materials-tarnish prediction—exploit precise, quantifiable features as indicators of trace presence, erasure, or transformation. In stylometry, the attacker’s attempt to erase textual “traces” may paradoxically induce even stronger forensic signals. In surface science, first-principles descriptors permit a priori identification of alloy systems resilient to corrosion-initiation, guiding both computational screening and experimental diagnostics. This convergence suggests a broader principle: any operation to erase or mask a trace, whether linguistic or electrochemical, leaves a quantifiable, and potentially amplified, secondary trace. Quantitative detection and mitigation thus remain an adversarial “race” driven by feature engineering and high-throughput analytics (Dilworth, 3 Dec 2025, Tiwari et al., 2021).