💻 computer science

Hybrid Anomaly Detection for Bullion Coin Authentication Leveraging Acoustic Signature Analysis

This paper proposes a hybrid deep learning framework that utilizes acoustic signature analysis and a dual-model architecture to achieve high-precision, non-destructive authentication of bullion coins while demonstrating potential scalability for critical component safety in automotive and aerospace industries.

Original authors: Krzysztof Siwek, Tran Hoai Linh, Tomasz Gryczka, Maciej Stodolski

Published 2026-05-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Krzysztof Siwek, Tran Hoai Linh, Tomasz Gryczka, Maciej Stodolski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a magic bell. When you tap a real gold or silver coin, it doesn't just make a "clink"; it sings a very specific, unique song. This song is determined by exactly what the coin is made of, how heavy it is, and its exact shape. If someone tries to fake the coin using cheap metal inside a shiny silver shell, the "song" changes. It might sound slightly off-key, or the notes might be muffled, even if the fake coin looks identical to the real one.

This paper describes a high-tech way to listen to that song and decide if a coin is real or a fake, without scratching or damaging it.

Here is how the researchers built their system, explained simply:

1. The Problem: The "Ear" Isn't Enough

For centuries, people have tapped coins to listen for a "ring." If it sounds like a dull thud, it's probably fake. But today's counterfeiters are very good. They can make fake coins that look perfect and even sound almost right to a human ear. Humans get tired, and their ears aren't precise enough to catch the tiny differences in high-quality fakes.

2. The Solution: The "Acoustic Fingerprint"

The researchers realized that every coin has a unique "acoustic fingerprint." Just as your fingerprint is unique to your hand, a coin's sound frequency is unique to its material.

The Real Coin: When struck, it vibrates at specific, predictable notes (like 3,770 Hz, 8,648 Hz, etc.).
The Fake Coin: Even if it looks the same, the fake (often made with a tungsten core) vibrates differently. Its notes are slightly shifted or missing.

3. The Brain: Two AI Models Working Together

The team built a computer system with two main parts, working like a security team:

Part A: The "Reconstruction Artist" (The Autoencoder)
Think of this as an artist who has only ever seen pictures of real coins.

Training: The artist studies thousands of photos of real coins and learns exactly what a "real" sound looks like.
The Test: When you show the artist a new coin's sound, they try to redraw it from memory.
- If it's real: The artist can redraw it perfectly because it matches what they've seen before. The "error" (the difference between the real sound and the drawing) is tiny.
- If it's fake: The artist gets confused. They try to force the fake sound to look like a real one, but it doesn't fit. The drawing comes out messy and wrong. The "error" is huge.
The Verdict: If the error is too big, the system screams, "This isn't real!" This is called Anomaly Detection. It's great because the system doesn't need to have seen the specific fake before; it just knows the fake doesn't look like the "real" pattern it learned.

Part B: The "Name Tagger" (The Classifier)
Once the system confirms a coin is likely real, a second AI model steps in. This one is like a librarian who knows the specific "names" of different coins (e.g., "Australian Kangaroo" vs. "Athenian Owl"). It looks at the unique notes of the real coin and tells you exactly which type it is.

4. The Challenges: Learning with Few Examples

The researchers faced a big hurdle: they didn't have a huge library of fake coins to study (because collecting verified fakes is hard and illegal in some places).

The Workaround: They used a trick called Data Augmentation. Imagine taking one recording of a coin and playing it slightly louder, slightly quieter, or adding a little bit of background static (like wind or traffic noise). This creates "new" versions of the same sound, teaching the AI to be tough against bad recording conditions.
The Result: Even with limited data, the system learned to spot the difference between a real coin and a fake one with high precision.

5. What Happened in the Tests?

The team tested their system with several scenarios:

Real Coin: The system recognized it immediately, matched the "song" perfectly, and identified the coin type.
The Tungsten Fake: They tested a fake coin made of tungsten covered in silver. Visually, it was indistinguishable from the real one. But when the AI "listened," the fake's song was so different that the "Reconstruction Artist" couldn't copy it. The error score was massive, and the system correctly flagged it as a fake.
The "Unknown" Real Coin: They tested a different real coin (a Vienna Philharmonic) that the AI had never seen before. Because the AI had only learned the "songs" of Kangaroo and Owl coins, it didn't recognize the Vienna coin's song. It flagged it as an anomaly (a fake).
- Why? The system is designed to be very safe. If it doesn't know the song, it assumes it's a fake rather than risking a mistake.

Summary

This paper presents a non-destructive way to authenticate bullion coins. By using AI to listen to the "song" a coin makes when tapped, the system can detect fakes that look perfect to the human eye. It works by learning what a "real" sound looks like and rejecting anything that doesn't fit that pattern. While the system is currently very good at spotting fakes, it is also very strict: if it hasn't been trained on a specific type of real coin, it will treat it as suspicious.

1. Problem Statement

The verification of bullion coin authenticity is critical for the integrity of the precious metals market. Traditional inspection methods (visual, weighing, displacement) are increasingly insufficient against sophisticated counterfeits, such as coins with tungsten cores plated in silver.

The Challenge: High-quality counterfeits are visually indistinguishable from genuine coins but possess different material properties (density, Young's modulus) that alter their acoustic resonance.
Data Scarcity: Developing machine learning models for this task is hindered by a lack of large, diverse datasets. Specifically, verified acoustic data for counterfeit coins is virtually non-existent due to legal and logistical constraints. This creates a "one-class" classification problem where only authentic samples are available for training.

2. Methodology

The authors propose a non-destructive verification framework combining acoustic frequency analysis with a hybrid deep learning architecture.

A. Data Acquisition and Preprocessing

Excitation: Coins are mechanically struck with a rigid object to induce natural vibrational modes.
Signal Processing:
- Segmentation: The recording is trimmed to isolate the impact and subsequent resonance, removing silence and excess noise.
- Normalization: Root Mean Square (RMS) normalization is applied to standardize loudness levels across different recording devices and distances.
- Augmentation: To combat data scarcity, the dataset is expanded using:
  - Amplitude scaling.
  - Addition of Gaussian noise (simulating environmental interference).
  - Hamming windowing to minimize spectral leakage.
Feature Extraction: Fast Fourier Transform (FFT) is used to generate frequency spectra. A peak detection algorithm identifies dominant resonance frequencies and amplitudes.

B. Hybrid Model Architecture

The system utilizes two synergistic models:

Autoencoder (Anomaly Detection):
- Type: Unsupervised Deep Autoencoder.
- Training: Trained exclusively on authentic coin spectra.
- Mechanism: The encoder compresses the input spectrum (8,820 features) into a latent space (128 neurons), and the decoder attempts to reconstruct the original spectrum.
- Logic: The model learns to perfectly reconstruct authentic coins (low reconstruction error). When presented with a counterfeit (or unknown coin), the model fails to reconstruct the spectrum accurately, resulting in a high reconstruction error.
- Thresholding: A dynamic anomaly threshold is calculated using the formula: $Threshold = \mu_d + 3\sigma_d$ , where $\mu_d$ and $\sigma_d$ are the mean and standard deviation of the peak distances between original and reconstructed spectra. This ensures 99.75% of authentic samples are accepted.
Classifier (Type Identification):
- Input: The latent representation (intermediate features) generated by the autoencoder's encoder.
- Function: A two-layer neural network classifies the coin type (e.g., Australian Kangaroo vs. Athenian Owl).
- Logic: If the anomaly detector flags a coin as "unknown," the classifier does not attempt to assign a type, preventing false positives on genuine but untrained coin types.

3. Key Contributions

One-Class Learning Solution: The paper addresses the critical lack of counterfeit training data by utilizing an autoencoder-based anomaly detection approach, allowing the system to identify anomalies without ever seeing a fake coin during training.
Dynamic Thresholding: Instead of a static error cutoff, the system calculates a statistical threshold based on the training distribution ( $\mu + 3\sigma$ ), making it robust against varying recording conditions.
Hybrid Synergy: The combination of an unsupervised anomaly detector and a supervised classifier allows the system to not only detect fakes but also identify specific coin types, while safely rejecting unknown genuine coins as "anomalies" (a conservative security posture).
Non-Destructive Testing (NDT): The method validates the use of acoustic signatures as a reliable, non-invasive physical fingerprint for material verification.

4. Experimental Results

The system was tested on 1 oz silver coins (Australian Kangaroo, Athenian Owl, Vienna Philharmonic) and a tungsten-core counterfeit.

Authentic Coin (Trained): The autoencoder perfectly reconstructed the spectrum. The peak distance was 0.13, well below the threshold of 131. The coin was correctly identified as authentic and classified as an Australian Kangaroo.
Counterfeit Coin (Tungsten Core): Due to the different material properties, the autoencoder failed to reconstruct the spectrum. The peak distance was 1077, far exceeding the threshold. The system correctly flagged it as a counterfeit (anomaly) and halted further classification.
Authentic Coin (Untrained): A genuine Vienna Philharmonic coin (not in the training set) was reconstructed poorly, yielding a distance of 1708.
- Result: The system correctly flagged it as an anomaly (counterfeit/unknown).
- Implication: This demonstrates the system's conservative security stance: it prioritizes rejecting unknowns over misclassifying them, though it highlights a limitation in recognizing new genuine coin types without retraining.

5. Significance and Future Directions

Market Impact: The study provides a scalable, low-cost, and rapid method for verifying bullion, addressing a major vulnerability in the precious metals market.
Broader Application: The methodology is not limited to coins; it is applicable to Non-Destructive Testing (NDT) for critical components in the automotive and aerospace industries where material integrity is paramount.
Limitations: The current model is sensitive to "novel" data (unseen genuine coin types are rejected). The authors note that the small dataset size is a primary constraint.
Future Work:
- Implementation of Deep Variational Autoencoders (VAE) for better probabilistic modeling.
- Exploration of "Tiny" Large Language Models (LLMs) for zero-shot and few-shot learning to reduce reliance on massive datasets.
- Expansion of the training dataset to include a wider variety of minting series.

In conclusion, the paper successfully demonstrates that a hybrid deep learning approach, leveraging acoustic signatures and unsupervised anomaly detection, can effectively distinguish high-quality counterfeits from genuine bullion coins, even in the absence of counterfeit training data.