Synthetic Media, Deepfakes, and Digital Trust

An evidence-based account of what fails, what holds, and how each claim can be checked

Author: Shahzad Asghar — MSc Computer Science, CISSP, CISM, PMP, PRINCE2, TOGAF
Version: 1.0, in progress
Evidence reviewed through: 24 October 2026
Publisher: ShahzadAsghar.com

Independence and scope

This book is the independent work of the author. It does not represent the position of the United Nations, of any employer, or of any standards body cited within it.

What this book is not. It is not a legal manual, and nothing in it is legal advice: regulatory status changes and must be confirmed with counsel in each jurisdiction. It is not a product or tool review, and no detection vendor is endorsed. It is not a guide to producing synthetic media, and it deliberately omits operational detail that would help anyone do so.

The argument this book tests

Synthetic media is not solely a problem of identifying manipulated pixels, voices, or text. It is a problem of identity, evidence, institutional trust, verification processes, and organizational decision-making.

Detection matters in some contexts and is close to useless in others. The uncomfortable finding that runs through the research literature is that the quality of a forgery is often irrelevant to whether an attack succeeds, because the forgery is rarely what fails. What fails is a payment process that treats a familiar face as an identity check, an evidence rule that assumes recordings are self-authenticating, or a hierarchy in which no one is permitted to question an instruction that appears to come from a senior executive.

That claim needs testing rather than asserting, and each chapter closes by mapping its findings onto the five-layer framework below. If the framework survives twelve chapters of contrary evidence, it is useful. If it does not, the book will say so.

Figure 0.1 — The five-layer trust framework used throughout this book. An attack does not need to defeat every layer, only to reach a decision before any layer stops it. Detection is drawn inside provenance deliberately: a detector answers "does this look generated?", while provenance answers "where did this come from?", and only the second survives an adversary who controls the artefact. Source: author's framework, developed in Chapter 12 and tested against the evidence in each chapter.

How to read this book

The twelve chapters run in four parts. Part I establishes how synthetic media is generated and what the documented attacks actually did: Foundations of Synthetic Media Generation, A Taxonomy of AI-Generated Disinformation Campaigns and Deepfakes as a Cybersecurity Threat Vector. Part II examines detection and is the longest sceptical stretch: Deep Learning and Multimodal Deepfake Detection, Voice Cloning and Audio Deepfake Detection, Benchmarking, Datasets and Cross-Domain Generalisation and Adversarial Robustness and Evasion. Part III covers text and what is arriving next: LLM-Generated and Multimodal Disinformation and Emerging Generative Threats. Part IV turns constructive: Content Provenance and Watermarking, The Legal and Regulatory Landscape and Building an Organisational Defence Framework.

A reader who wants the practical answer rather than the argument can start at Chapter 12, which ranks nine controls by the evidence behind each one, then read Deepfakes as a Cybersecurity Threat Vector for the cases those rankings come from.

Every substantial factual claim carries an identifier resolving to the evidence ledger. The full contents lists all twelve chapters with reading times.