Obfuscation of Audio Signals

Principal Investigator: Hany Abdel-Khalik

The current industrial paradigm has seen an increased variety in data volume, real-time analytics, and the role of artificial intelligence and machine learning (AI/ML) tools that process many types of data, including timeseries, image, text-based, and audio signals to inform operational decisions. While the corresponding rise in data utility has allowed for efficient, safe, and cheap data analytics, these tools often invite major security concerns in light of the increased risk of intrusion detection and data leakage. Among the chief security challenges is that of audio-based signals; whether this be machine-operation signals, proprietary conversations, etc., audio data may overtly detail sensitive information that necessitates a specialized form of obfuscation to ensure are protected in real-time.

In the case of data sharing, security is typically ensured by encryption-based techniques which require the genuine audio signal to be released. The so-called issue of trust can be solved by obfuscating the real signal with an arbitrary signal or image such that its identity is no longer at risk, thereby mitigating the uncertainties associated with typical collaboration. The Directed Infusion of Data (DIOD) paradigm can be directly applied to audio-based data, thus allowing a typical conversation to take the form of a song, another conversation, or an arbitrary signal.

A secondary benefit of the DIOD paradigm applied to audio-based data is that direct information of the Fourier-type analyses, i.e., the most fundamental frequencies of the audio data, or time-stamped signal properties, may be preserved to suit the needs of a downstream analysis, thus allowing covert data masking of full conversations while retaining characteristic attributes.

Personnel

Students: Chloe Yoder Tyler Lewis Arvind Sundaram

Keywords: data security audio embedding, information obfuscation