You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: docs/_posts/2026-02-06-Video-&-Audio-Provenance-In-The-Age-of-AI.md
+76-64Lines changed: 76 additions & 64 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -1,6 +1,6 @@
1
1
---
2
2
title: Video & Audio Provenance on Arm in the Age of AI
3
-
description: Integrating transparent provenance - disclosing whether media is AI-generated or AI-edited, and also what other AI processing has occurred on any media - is necessary for accountability in domains like journalism, security, and regulated industries. This project uses C2PA 2.3 to record such actions as signed, machine-verifiable assertions attached to the asset. This reduces deepfake proliferation and fraud by providing a method to verify provenance. Additionally it tracks authorised and potential unauthorised use of models to analyse the asset.
3
+
description: Integrating transparent provenance - disclosing whether media is AI-generated or AI-edited, and what other AI processing has occurred on any media - is fundamental for accountability in domains like journalism, security, and regulated industries. This project uses C2PA specification (www.c2pa.org) revision 2.3 to record such actions as signed, machine-verifiable assertions attached to the asset. This reduces deepfake proliferation and fraud by providing a method to verify provenance. Additionally, it tracks authorized and potential unauthorized use of models to analyze the asset, by means of Model Provenance.
4
4
subjects:
5
5
- ML
6
6
- Security
@@ -18,7 +18,7 @@ support-level:
18
18
publication-date: 2026-02-06
19
19
license:
20
20
status:
21
-
- Hidden
21
+
- Published
22
22
badges:
23
23
- trending
24
24
- recently_added
@@ -27,59 +27,65 @@ layout: article
27
27
sidebar:
28
28
nav: projects
29
29
full_description: |-
30
-
## Description
30
+
## Description
31
31
32
-
### Why is this important?
32
+
### Why is this important?
33
33
34
-
Generative AI has created a trust gap:
35
-
- Images, videos, and audio may be AI-generated or heavily manipulated - this can lead to deepfakes, and fraud.
36
-
- This AI-enhancement or editing can occur in real-time - i.e. a person can livestream themeself, but use ML to change their appearance and voice.
37
-
- Moreover, the use of AI models to scan images, video, and audio (e.g. facial or voice recognition) is often opaque, with no actual changes to the underlying data.
38
-
- Viewers rarely know whether AI analysis was performed, or what conclusion was reached.
34
+
Generative AI has created a trust gap:
35
+
- Images, videos, and audio may be AI-generated or heavily manipulated
36
+
- this can lead to deepfakes, and fraud.
37
+
- This AI-enhancement or editing can occur in real-time - i.e. a person can live stream themselves but use ML to change their appearance and voice.
38
+
- Moreover, the use of AI models to scan images, video, and audio (e.g. facial or voice recognition) is often opaque, with no actual changes to the underlying data.
39
+
- Viewers rarely know whether AI analysis was performed, or what conclusion was reached. Such seamless content transformation, does not require much specialized skills and hence within the reach of common masses.
39
40
40
-
Integrating transparent provenance - disclosing whether media is AI-generated or AI-edited, and also what other AI processing has occurred on any media - is necessary for accountability in domains like journalism, security, and regulated industries. This project uses C2PA 2.3 to record such actions as signed, machine-verifiable assertions attached to the asset. This reduces deepfake proliferation and fraud by providing a method to verify provenance. Additionally it tracks authorised and potential unauthorised use of models to analyse the asset.
41
+
Integrating transparent provenance - disclosing whether media is AI-generated or AI-edited, and what other AI processing has occurred on any media - is fundamental for accountability in domains like journalism, security, and regulated industries. This project uses C2PA specification (www.c2pa.org) revision 2.3 to record such actions as signed, machine-verifiable assertions attached to the asset. This reduces deepfake proliferation and fraud by providing a method to verify provenance. Additionally, it tracks authorized and potential unauthorized use of models to analyze the asset, by means of Model Provenance.
41
42
42
43
[Arm is a founder member of the C2PA Standards Group.](https://newsroom.arm.com/blog/c2pa-fights-disinformation) The Coalition for Content Provenance and Authenticity (C2PA) specification lets creators and platforms attach cryptographically signed metadata to content like:
43
44
- images 📷
44
45
- videos 🎥
45
46
- audio 🎧
46
47
- documents 📄
47
48
48
-
That metadata can include:
49
-
-who created it
50
-
-what tool was used (camera, Photoshop, generative AI, etc.)
51
-
-what edits were made and when
52
-
-whether AI was involved
49
+
The metadata can include:
50
+
-who created it
51
+
-what tool was used (camera, Photoshop, generative AI, etc.)
52
+
-what edits were made and when
53
+
-whether AI was involved
53
54
54
-
Anyone can later verify this info to see if the content is authentic or has been tampered with.
55
+
Anyone can later verify this info to see if the content is authentic or has been tampered with.
55
56
56
-
The best place to start would be with processing on static audio or video files, but we encourage you to ultimately target a streamed data format - i.e. streamed audio or video, and to perform the inference and record the actions in real-time.
57
+
The best place to start would be with processing on static audio or video files, but we encourage you to ultimately target a streamed data format - i.e. streamed audio or video, and to perform the inference and record the actions in real-time, via attaching provenance information on the streaming content.
57
58
58
-
You should leverage an Arm-powered device, such as a Windows-on-Arm laptop, Apple Silicon MacBook, Arm-powered mobile phone, or Raspberry Pi.
59
+
You should leverage an Arm-powered device, such as a Windows-on-Arm laptop, Apple Silicon MacBook, Arm-powered mobile phone, or Raspberry Pi.
59
60
61
+
### Project Summary
60
62
61
-
### Project Summary
63
+
Build and evaluate a comprehensive AI-augmented audio/video capture and provenance system on an Arm-powered device (e.g. Windows-on-Arm laptop)
64
+
- captures streamed media with a camera or microphone,
65
+
- runs AI models on-device (i.e. face/object/keyword/sentiment detection, upscaling/filters/enhancements),
66
+
- generates C2PA Content Credentials that transparently disclose:
67
+
1. which models were run,
68
+
2. their effect/impact on the image or video
62
69
63
-
Build and evaluate a comprehensive AI-augmented audio/video capture and provenance system on an Arm-powered device (e.g. Windows-on-Arm laptop)
64
-
- captures streamed media with a camera or microphone,
65
-
- runs AI models on-device (i.e. face/object/keyword/sentiment detection, upscaling/filters/enhancements),
66
-
- generates C2PA Content Credentials that transparently disclose:
67
-
1. which models were run,
68
-
2. their effect/impact on the image or video
70
+
and demonstrates how this provenance enables trust and auditability in real-world use cases such as content integrity validation and responsible media pipelines.
69
71
70
-
and demonstrates how this provenance enables trust and auditability in real-world use cases such as content integrity validation and responsible media pipelines.
72
+
You should be able to show your source code, along with documentation/instructions/comments, a short demo video or images to show the project in action, and a short document describing your decisions and how you implemented the project.
71
73
72
-
You should be able to show your source code, along with documentation/instructions/comments, a short demo video or images to show the project in action, and a short document describing your decisions and how you implemented the project.
74
+
## What will you use?
73
75
76
+
You should be willing to learn about, or already familiar with:
77
+
- Programming and building a basic application on your chosen platform/OS
78
+
- C2PA and the concepts behind content provenance and authentication
79
+
- Deploying optimized inference models on Arm-powered CPU via frameworks with KleidiAI integration (e.g. PyTorch).
74
80
75
-
## What will you use?
76
-
You should be willing to learn about, or already familiar with:
77
-
- Programming and building a basic application on your chosen platform/OS
78
-
- C2PA and the concepts behind content provenance and authentication
79
-
- Deploying optimised inference models on Arm-powered CPU via frameworks with KleidiAI integration (e.g. PyTorch)
81
+
82
+
## A few helpful links to relevant items:
83
+
- [Live Video Streaming](https://spec.c2pa.org/specifications/specifications/2.3/specs/C2PA_Specification.html#live-video)
84
+
- [C2PA](https://github.com/contentauth/c2pa-rs)
85
+
- [C-Wrapper for Rust Library](https://gitlab.com/guardianproject/proofmode/simple-c2pa)
80
86
81
87
82
-
## Resources from Arm and our partners
88
+
## Other potentially useful resources from Arm and our partners
83
89
- Repository: [AI on Arm course](https://github.com/arm-university/AI-on-Arm)
84
90
- Arm / Cambridge University edX course: [AI at the Edge on Arm (Mobile)](https://www.edx.org/learn/computer-science/arm-education-ai-at-the-edge-on-arm)
85
91
- Learning Path: [Vision LLM Inference on Android with KleidiAI](https://learn.arm.com/learning-paths/mobile-graphics-and-gaming/vision-llm-inference-on-android-with-kleidiai-and-mnn/)
@@ -100,59 +106,65 @@ full_description: |-
100
106
To receive the benefits, you must show us your project through our [online form](https://forms.office.com/e/VZnJQLeRhD). Please do not include any confidential information in your contribution. Additionally if you are affiliated with an academic institution, please ensure you have the right to share your material.
101
107
---
102
108
103
-
## Description
109
+
## Description
104
110
105
-
### Why is this important?
111
+
### Why is this important?
106
112
107
-
Generative AI has created a trust gap:
108
-
- Images, videos, and audio may be AI-generated or heavily manipulated - this can lead to deepfakes, and fraud.
109
-
- This AI-enhancement or editing can occur in real-time - i.e. a person can livestream themeself, but use ML to change their appearance and voice.
110
-
- Moreover, the use of AI models to scan images, video, and audio (e.g. facial or voice recognition) is often opaque, with no actual changes to the underlying data.
111
-
- Viewers rarely know whether AI analysis was performed, or what conclusion was reached.
113
+
Generative AI has created a trust gap:
114
+
- Images, videos, and audio may be AI-generated or heavily manipulated
115
+
- this can lead to deepfakes, and fraud.
116
+
- This AI-enhancement or editing can occur in real-time - i.e. a person can live stream themselves but use ML to change their appearance and voice.
117
+
- Moreover, the use of AI models to scan images, video, and audio (e.g. facial or voice recognition) is often opaque, with no actual changes to the underlying data.
118
+
- Viewers rarely know whether AI analysis was performed, or what conclusion was reached. Such seamless content transformation, does not require much specialized skills and hence within the reach of common masses.
112
119
113
-
Integrating transparent provenance - disclosing whether media is AI-generated or AI-edited, and also what other AI processing has occurred on any media - is necessary for accountability in domains like journalism, security, and regulated industries. This project uses C2PA 2.3 to record such actions as signed, machine-verifiable assertions attached to the asset. This reduces deepfake proliferation and fraud by providing a method to verify provenance. Additionally it tracks authorised and potential unauthorised use of models to analyse the asset.
120
+
Integrating transparent provenance - disclosing whether media is AI-generated or AI-edited, and what other AI processing has occurred on any media - is fundamental for accountability in domains like journalism, security, and regulated industries. This project uses C2PA specification (www.c2pa.org) revision 2.3 to record such actions as signed, machine-verifiable assertions attached to the asset. This reduces deepfake proliferation and fraud by providing a method to verify provenance. Additionally, it tracks authorized and potential unauthorized use of models to analyze the asset, by means of Model Provenance.
114
121
115
122
[Arm is a founder member of the C2PA Standards Group.](https://newsroom.arm.com/blog/c2pa-fights-disinformation) The Coalition for Content Provenance and Authenticity (C2PA) specification lets creators and platforms attach cryptographically signed metadata to content like:
116
123
- images 📷
117
124
- videos 🎥
118
125
- audio 🎧
119
126
- documents 📄
120
127
121
-
That metadata can include:
122
-
-who created it
123
-
-what tool was used (camera, Photoshop, generative AI, etc.)
124
-
-what edits were made and when
125
-
-whether AI was involved
128
+
The metadata can include:
129
+
-who created it
130
+
-what tool was used (camera, Photoshop, generative AI, etc.)
131
+
-what edits were made and when
132
+
-whether AI was involved
126
133
127
-
Anyone can later verify this info to see if the content is authentic or has been tampered with.
134
+
Anyone can later verify this info to see if the content is authentic or has been tampered with.
128
135
129
-
The best place to start would be with processing on static audio or video files, but we encourage you to ultimately target a streamed data format - i.e. streamed audio or video, and to perform the inference and record the actions in real-time.
136
+
The best place to start would be with processing on static audio or video files, but we encourage you to ultimately target a streamed data format - i.e. streamed audio or video, and to perform the inference and record the actions in real-time, via attaching provenance information on the streaming content.
130
137
131
-
You should leverage an Arm-powered device, such as a Windows-on-Arm laptop, Apple Silicon MacBook, Arm-powered mobile phone, or Raspberry Pi.
138
+
You should leverage an Arm-powered device, such as a Windows-on-Arm laptop, Apple Silicon MacBook, Arm-powered mobile phone, or Raspberry Pi.
132
139
140
+
### Project Summary
133
141
134
-
### Project Summary
142
+
Build and evaluate a comprehensive AI-augmented audio/video capture and provenance system on an Arm-powered device (e.g. Windows-on-Arm laptop)
143
+
- captures streamed media with a camera or microphone,
144
+
- runs AI models on-device (i.e. face/object/keyword/sentiment detection, upscaling/filters/enhancements),
145
+
- generates C2PA Content Credentials that transparently disclose:
146
+
1. which models were run,
147
+
2. their effect/impact on the image or video
135
148
136
-
Build and evaluate a comprehensive AI-augmented audio/video capture and provenance system on an Arm-powered device (e.g. Windows-on-Arm laptop)
137
-
- captures streamed media with a camera or microphone,
138
-
- runs AI models on-device (i.e. face/object/keyword/sentiment detection, upscaling/filters/enhancements),
139
-
- generates C2PA Content Credentials that transparently disclose:
140
-
1. which models were run,
141
-
2. their effect/impact on the image or video
149
+
and demonstrates how this provenance enables trust and auditability in real-world use cases such as content integrity validation and responsible media pipelines.
142
150
143
-
and demonstrates how this provenance enables trust and auditability in real-world use cases such as content integrity validation and responsible media pipelines.
151
+
You should be able to show your source code, along with documentation/instructions/comments, a short demo video or images to show the project in action, and a short document describing your decisions and how you implemented the project.
144
152
145
-
You should be able to show your source code, along with documentation/instructions/comments, a short demo video or images to show the project in action, and a short document describing your decisions and how you implemented the project.
153
+
## What will you use?
146
154
155
+
You should be willing to learn about, or already familiar with:
156
+
- Programming and building a basic application on your chosen platform/OS
157
+
- C2PA and the concepts behind content provenance and authentication
158
+
- Deploying optimized inference models on Arm-powered CPU via frameworks with KleidiAI integration (e.g. PyTorch).
147
159
148
-
## What will you use?
149
-
You should be willing to learn about, or already familiar with:
150
-
-Programming and building a basic application on your chosen platform/OS
151
-
- C2PA and the concepts behind content provenance and authentication
152
-
-Deploying optimised inference models on Arm-powered CPU via frameworks with KleidiAI integration (e.g. PyTorch)
160
+
161
+
## A few helpful links to relevant items:
162
+
-[Live Video Streaming](https://spec.c2pa.org/specifications/specifications/2.3/specs/C2PA_Specification.html#live-video)
163
+
-[C2PA](https://github.com/contentauth/c2pa-rs)
164
+
-[C-Wrapper for Rust Library](https://gitlab.com/guardianproject/proofmode/simple-c2pa)
153
165
154
166
155
-
## Resources from Arm and our partners
167
+
## Other potentially useful resources from Arm and our partners
156
168
- Repository: [AI on Arm course](https://github.com/arm-university/AI-on-Arm)
157
169
- Arm / Cambridge University edX course: [AI at the Edge on Arm (Mobile)](https://www.edx.org/learn/computer-science/arm-education-ai-at-the-edge-on-arm)
158
170
- Learning Path: [Vision LLM Inference on Android with KleidiAI](https://learn.arm.com/learning-paths/mobile-graphics-and-gaming/vision-llm-inference-on-android-with-kleidiai-and-mnn/)
0 commit comments