LenslessMic: Optical Audio Encryption

• LenslessMic is an optical encryption method that converts audio data into secure visual representations using a lensless camera with a programmable mask.
• The approach integrates a neural audio codec with a lensless imaging system to map audio signals into images, where the physical PSF serves as the encryption key.
• Experimental validation with a Raspberry Pi prototype demonstrates that precise PSF knowledge enables robust decryption, achieving security levels comparable to AES-256.

LenslessMic is an optical encryption and authentication methodology that leverages lensless computational imaging to provide a physical-layer of security for audio data. Unlike traditional audio encryption, which is implemented purely in the digital or signal-processing domains, LenslessMic uses a physical optical system—specifically, a lensless camera with a programmable mask—to multiplex audio representations as images, creating a robust coupling between the security level and the physical system parameters. The approach is validated with a low-cost Raspberry Pi–based prototype, and open-source datasets are provided to enable further research and application development (Grinberg et al., 19 Sep 2025).

1. Principle of Operation: Optical–Digital Hybrid for Audio Encryption

LenslessMic achieves encryption by transforming audio data into a visual domain and encoding it optically via lensless computational imaging. The key process consists of:

• Audio-to-Visual Mapping: The raw audio signal is compressed with a neural audio codec (NAC), typically one that produces latent embeddings (such as the DAC codec). The latent variable is reshaped into a compact time-varying image (or sequence of images), denoted as $V$.
• Optical Multiplexing: The visual representation $V$ is displayed on a screen and then optically recorded by a lensless camera. The camera employs a programmable amplitude (or phase) mask, which modulates the incoming light before it reaches the sensor, encoding the image via a point-spread function (PSF).
• Forward Model: The sensor measurement $y$ is modeled as

$y = Hx + n$

where $x$ is the vectorized visual representation of the audio, $H$ is the system matrix determined by the PSF and the physical mask, and $n$ is sensor noise.

The crucial security property is the dependence of successful decryption (recovery of $x$ from $y$) on knowledge of the exact PSF (mask). Even minute deviations in $H$ (for example, incorrect mask bits) drastically reduce the quality of the reconstructed audio.

2. Hardware Architecture and the Role of the Lensless Camera

The LenslessMic realization is based on an inexpensive hardware setup that integrates:

• Lensless Camera Prototype: Built on a Raspberry Pi platform with a programmable mask, typically implemented via an amplitude-modulating liquid crystal or binary optical mask. The mask is placed directly in front of the sensor, removing all lenses from the optical path.
• Programmable Mask as Key: The mask’s pattern directly defines the PSF, which acts as the encryption key. Changing the mask changes the mapping $H$, thus making the encryption physically reconfigurable.
• Optical Capture: The audio-derived image $V$ is displayed on a monitor and “imaged” by the lensless camera. The resulting sensor data is a multiplexed, unintelligible snapshot bearing no direct resemblance to the input image or the original audio.

In the decryption (reconstruction) phase, access to the correct PSF (i.e., knowledge of $H$) is needed to solve the inverse problem and recover $V$, enabling audio decoding via the neural codec.

3. Mathematical Foundations of Security and Authentication

The encryption and authentication strength in LenslessMic derive from the complexity of the programmable mask and the sensitivity of decryption to correct PSF knowledge.

• System Matrix Properties: $H$ is generally a block-circulant (or block-structured) matrix determined by the mask pattern, with each column a shifted instance of the PSF.
• Authentication: Users are associated with unique mask patterns $P_u$, and recordings are tagged accordingly. Correct decryption (authentication) requires supplying the exact $P_u$; otherwise, the output quality collapses.
• Decryption Error Sensitivity:

$$\hat{x} = \hat{H}^{-1} y = x + H^{-1} n + [ (x + H^{-1} n) \cdot \sum_{k=1}^{\infty} (H^{-1}\Delta)^k ]$$

where $\Delta = H - \hat{H}$ is the mismatch in the system matrix. Even small errors in $\hat{H}$ exponentially degrade reconstruction due to powers of $H^{-1}\Delta$.

• Encryption Strength and Search Space:

For a programmable mask with $N$ independently determined pixels and $b$ possible intensity levels, the total number of possible keys is $b^N$. The required minimum correct PSF ratio $W$ for $K$-bit security is given by:

$W \geq \frac{K \log_b 2}{N}$

As shown in the implementation, $N = 1296$, $b = 8$, so that $W \approx 7\%$ ensures AES-256 level brute-force security.

• Robustness to Chosen/Known-Plaintext Attacks (CPA/KPA): By grouping multiple frames (i.e., increasing the time window and super-pixel grouping), the system increases attack difficulty since the attacker must simultaneously guess multiple PSF subregions.

4. Audio Reconstruction and Codec Integration

LenslessMic employs advanced audio compression and error-robust reconstruction pipelines:

• Neural Audio Codec: The latent representation is highly entropic and noise-like (due to residual vector quantization, RVQ). Decoding the latent requires matching codebook indices, which is robust to moderate noise but highly sensitive to PSF errors.
• Reconstruction Algorithms: Inversion is performed via computational imaging solvers, notably ADMM (Alternating Direction Method of Multipliers) unrolled in a learned fashion—often incorporating DRUNet-style (denoising residual UNet) modules for denoising and pre/post-PSF correction.
• Super-Pixel Downsampling: For resilience, raw measurements are pooled into “super-pixels” (by projecting $V$ to lower resolution before encoding). This aids error correction and tolerates small reconstruction imperfections while still requiring the correct PSF for content integrity.

5. Experimental Validation and Open-Source Resources

Experimental results are based on:

• Low-Cost Raspberry Pi Prototype: Total cost approximately $100 USD, validating LenslessMic’s accessibility.
• Data Modalities: Speech (Librispeech) and music (SongDescriber) were used as test cases, encoded first via NAC and then subjected to optical encryption.
• Authentication Evaluation: The system was tested with various mask keys, verifying that even small PSF mismatches (or bit errors) led to unintelligible audio (e.g., word error rate/WER of 100%). Only the exact key yielded low WER and high UT-MOS (a quality metric).
• Open-Source Datasets: The creators released datasets comprising tens of thousands of lensless encrypted frames, supporting further research and benchmarking in the community.

6. Applications, Limitations, and Security–Quality Trade-Offs

Potential Applications

• Physical-Layer Secure Audio Communication: Prevent unauthorized interception or eavesdropping at the hardware level.
• Dual Encryption and Authentication: Ensure only users with the correct PSF can decrypt and authenticate audio messages.
• Multi-Modal Support: Applicable to voice, music, and other complex audio via the neural audio codec interface.
• Hardware Key Reconfiguration: The programmable mask enables frequent key changes, addressing long-term security needs.

Limitations

• Frame Rate and Latency: Optical capture is slower than real-time audio; mitigation relies on chunked audio and codec compression.
• Storage Overhead: Storing lensless measurements may require more space than compressed digital audio.
• Physical Bulk: The prototype’s use of a monitor and lensless camera is larger than standard microphones; miniaturization would be needed for wide deployment.
• Security–Quality Trade-Off: Higher security (requiring more PSF knowledge or grouping multiple frames) can degrade audio quality or increase decoding latency.

7. Quantitative Security and Decryption Formula Table

Parameter	Symbol	Description
PSF pixels	N	Number of programmable mask elements
Bit depth	b	Number of intensity levels per PSF pixel
Security level (bits)	K	Target key strength (e.g., 256 for AES-256 equivalence)
Fraction needed	W	Min. correct PSF knowledge for decryption
PSF search space	$b^N$	Total possible mask patterns
Correctness formula	$W\ge\frac{K \log_b 2}{N}$	PSF correctness for K-bit strength
Decryption formula	$$\hat{x} = \hat{H}^{-1}y = x + H^{-1}n + [(x + H^{-1}n) \cdot \sum_{k=1}^{\infty}(H^{-1}\Delta)^k]$$	Error propagation when PSF is incorrect

The system’s security derives from the structural randomness of the PSF, tunable via mask complexity ($N, b$), and the exponential sensitivity of reconstruction to PSF errors.

Conclusion

LenslessMic exploits the physical inseparability of lensless computational imaging to create a new hybrid optical–digital encryption paradigm for audio. Encryption strength is mathematically tied to mask complexity, and authentication is enforced by strict dependence on knowledge of the PSF. The method is validated experimentally on speech and music, demonstrates security equivalent to modern cryptographic standards for appropriate PSF sizes, and is supported by an open-source infrastructure. Future developments may target miniaturization, increased throughput, and integration with other modalities for secure multi-modal sensing (Grinberg et al., 19 Sep 2025).