Amnestic Forgery: Semantics & Forensics

• Amnestic forgery is the systematic elimination of diagnostic evidence, merging metaphor blending in semantics with adversarial techniques in digital forensics.
• In digital forensics, strategies like SEAR and ForgeryEraser remove pixel-level and feature-space signatures, achieving high attack rates while preserving image quality.
• In ontological modeling, it integrates conceptual metaphors via structured frame mappings, advancing NLP applications and referential analysis.

Amnestic forgery denotes interventions—conceptual or computational—that erase or conceal all traces, statistical or semantic, of an originating entity or event. The term appears in distinct technical domains: (1) as the ontological modeling of metaphorical mappings with referential "blending" in natural language semantics (Gangemi et al., 2018); (2) as adversarial attacks on digital media forensics, where the objective is to eliminate any learnable signature of manipulation such that detectors (global or pixel-level) have "amnesia" regarding the tampering event (Zhuo et al., 2023, Li et al., 6 Feb 2026). Across these fields, amnestic forgery is characterized by the systematic removal of diagnostic evidence, rendering detection and attribution fundamentally infeasible.

1. Ontological Foundations: Amnestic Forgery as Metaphor Blending

The Amnestic Forgery (AF) ontology (Gangemi et al., 2018) systematizes conceptual metaphors by embedding MetaNet’s repository within the Framester schema. AF represents metaphors as Description entities in a D&S (Descriptions & Situations) pattern, and operationalizes mappings as partial functions between roles (FrameElements) of source and target frames: $m = (\,F_s,\,F_t,\,\pi\,)$ where $F_s$ and $F_t$ are source and target frames, and $\pi: FE(F_s) \rightharpoonup FE(F_t)$.

Blending, a key concept, creates a new frame $F_b = \text{Blend}(F_s,F_t,\pi)$, whose set of roles is defined as: $$FE(F_b) = (FE(F_s)\setminus\mathrm{dom}(\pi)) \cup (FE(F_t)\setminus\mathrm{cod}(\pi)) \cup \pi(\mathrm{dom}(\pi)).$$ This mechanism models referential "amnestic" effects: in metaphors such as CRIME IS A DISEASE, a "blended" CrimeDiseaseEvent inherits referents by aligning properties of the source and target, erasing extensional traces of the original conceptual divide.

2. Amnestic Forgery in Digital Forensics: Motivation and Definitions

In image anti-forensics, "amnestic forgery" references memory-erasing attacks that erase all statistical or machine-learned indicators of tampering, surpassing simple detector evasion. Traditional adversarial or manipulation operations (e.g., JPEG smoothing) only flip labels for entire images or induce easily-recognized zero-one reversals at the pixel level. By contrast, true amnestic attacks must ensure that forensic detection models (global classifiers or pixel-wise localizers) cannot distinguish forgeries from pristine data by any measurable artifact (Zhuo et al., 2023, Li et al., 6 Feb 2026).

This necessity is accentuated by the shift from binary tamper detection to pixel-level localization (SATFL, Mantra-Net, SPAN), where spatially precise concealment is required to defeat state-of-the-art forensics.

3. Algorithmic Approaches: SEAR and ForgeryEraser

Two principal strategies implement amnestic forgeries in computational image forensics:

SEAR (Self-supErvised Anti-foRensics) (Zhuo et al., 2023):

• Trains a concealer network $G$ to predict perturbations $\delta$ such that the perturbed image $I_{anti} = I_{forgery} + \delta$ fools any forgery localization network $F_{loc}$.
• Stage 1 uses a self-supervised pretext task to ensure perturbations are applied only at ground-truth tampered locations (mask $F_s$0), formally:

$F_s$1

• Stage 2 employs adversarial training, maximizing the cross-entropy between predicted and true masks:

$F_s$2

• The total loss for updating $F_s$3 is $F_s$4.

ForgeryEraser (Li et al., 6 Feb 2026):

• Implements a universal, black-box anti-forensics attack by manipulating image embeddings within a shared Vision-LLM (VLM; e.g., CLIP).
• Drives embeddings of forged images toward "authentic" text-derived anchors and repels them from "forgery" anchors using a multi-modal guidance loss:

$F_s$5

where

$F_s$6

• Utilizes a Momentum Iterative FGSM optimizer under an $F_s$7 perturbation budget.

Both frameworks operationalize "amnestic" attacks at different abstraction levels: SEAR conceals pixel-level artifacts, while ForgeryEraser targets feature-space memory in downstream detector backbones.

4. Formal Schema and Querying: Ontological and Computational Integration

In ontological terms, AF’s schema unifies metaphoric and referential erasure via OWL patterns:

• Classes: Metaphor, MetaphoricRoleMapping.
• Object properties: hasSourceFrame, hasTargetFrame, hasRoleMapping, mapsFrom, mapsTo. Example SPARQL queries against the Framester endpoint illustrate schema operation:
• Enumerate metaphors and frames
• Retrieve role mappings for a metaphor
• Identify adjective-noun synset pairs as candidates for metaphor generation

These mechanisms showcase AF’s ability to support metaphor generation, guide referential analysis, and interface with deep NLP pipelines for conceptual disambiguation (Gangemi et al., 2018).

In computational forensics, ForgeryEraser and SEAR are evaluated via metrics such as pixel-level F1, Attack Rate, SSIM, and PSNR; both have reproducible architectures described in detail, including input pipelines, loss hyperparameters, and adversarial optimization schemes (Zhuo et al., 2023, Li et al., 6 Feb 2026).

5. Empirical Findings and Efficacy

SEAR achieves attack rates up to 0.9969 in white-box settings, with virtually no zero-one mask reversal ($F_s$8), outperforming adversarial baselines by large margins. It transfers effectively to black-box scenarios and retrained forensic defenses, maintaining high SSIM (0.9816–0.9982) and PSNR (38–54 dB) (Zhuo et al., 2023). SEAR's inference speed reaches 0.16 s/image, exceeding classic adversarial methods.

ForgeryEraser causes relative accuracy drops exceeding 85% at standard budgets ($F_s$9) across six detector architectures, with certain detectors reduced to 1% accuracy. Feature-space analyses (t-SNE) show adversarial forgeries merging into real clusters, and explanation models supplying "authentic" rationales post-attack. Attack robustness persists under JPEG compression and Gaussian blur (Li et al., 6 Feb 2026).

These results confirm the practicality and generalizability of amnestic attacks over a broad set of detectors and manipulation types.

6. Applications and Broader Implications

In ontological linguistics, AF enables:

• Large-scale metaphor generation via lexical alignments (e.g., "forgery is amnesia" → "amnesic forgery"),
• Referential analysis of blended situations, supporting the study of quasi-truth in discourse,
• Integration into NLP pipelines for improved sense disambiguation and inferencing (Gangemi et al., 2018).

In image anti-forensics, amnestic attacks undermine current forensic architectures:

• SEAR demonstrates that concealers requiring only the forged image at inference can generalize across both black-box and retrained detectors,
• ForgeryEraser reveals systemic vulnerability—feature transferability via shared public VLM backbones—leading to universal evasion and erasure of forensic "memory" (Li et al., 6 Feb 2026, Zhuo et al., 2023).

A plausible implication is an ongoing "arms-race": purely data-driven or backbone-dependent detection models will remain inherently susceptible unless augmented by stronger priors or anomaly mechanisms beyond adversarial reach.

7. Open Challenges and Limitations

SEAR requires access to ground-truth tamper masks during training, which may not be available to real-world adversaries. This suggests the necessity of semi- or unsupervised methods for broader deployment. While image quality often remains imperceptibly affected, minute smoothing artifacts can occasionally betray prior tampering to advanced or hybrid forensic approaches. Perceptual discrimination modules may improve robustness against such visual deviations (Zhuo et al., 2023). Both frameworks underscore the need for defensive strategies not exclusively reliant on highly transferable deep features or public DNN backbones (Li et al., 6 Feb 2026).

In cognitive modeling, AF’s approach raises questions about the full extensional blending of referents in metaphor—can real-world social or legal distinctions ever be completely erased, or only re-conceptualized? Continued ontological refinement and logical modeling are needed to delimit the boundaries of true "amnestic" mapping in both semantics and computation.