WEBVTT
Kind: captions
Language: en

1
00:00:00.040 --> 00:00:04.240
The alternating current running through these
power lines changes direction constantly.

2
00:00:04.240 --> 00:00:07.080
A full cycle back and forth,
50 times a second.

3
00:00:07.080 --> 00:00:10.400
60 in some of the parts of the world,
but here in Europe, 50 times a second.

4
00:00:10.400 --> 00:00:12.360
But that transmission
isn't perfectly efficient. 

5
00:00:12.360 --> 00:00:14.680
There are imperfections
and stray electromagnetic fields.

6
00:00:14.680 --> 00:00:15.960
And a little bit of that power 

7
00:00:15.960 --> 00:00:19.880
becomes vibration, which makes sound
at that frequency of 50 hertz.

8
00:00:19.880 --> 00:00:22.200
Or 100, or other multiples of 50.

9
00:00:22.200 --> 00:00:25.200
That's the "mains hum"
you can sometimes hear coming

10
00:00:25.200 --> 00:00:29.000
from power lines, or maybe from speakers
that aren't quite plugged in properly.

11
00:00:29.000 --> 00:00:32.520
Also, that hum isn't
exactly 50 cycles a second.

12
00:00:32.520 --> 00:00:35.520
The grid frequency shifts
slightly and constantly over time

13
00:00:35.520 --> 00:00:37.880
as the engineers in charge
balance supply and demand.

14
00:00:37.880 --> 00:00:40.120
It's only a tiny fraction of a shift,
but it's there.

15
00:00:40.120 --> 00:00:44.520
The whole grid runs at the same,
slightly wobbling, almost-50 hertz,

16
00:00:44.520 --> 00:00:46.000
no matter where you are. 

17
00:00:46.000 --> 00:00:49.520
That mains hum can be really annoying
for people recording audio.

18
00:00:49.520 --> 00:00:51.600
You can hear it in the background
sometimes. 

19
00:00:51.600 --> 00:00:54.560
But if you were to log
what that wobbling grid

20
00:00:54.560 --> 00:00:58.320
frequency was, every second of every day,
keep all that in a database,

21
00:00:58.320 --> 00:01:01.880
and then compare it to recordings
with mains hum on them,

22
00:01:01.880 --> 00:01:04.560
then in theory,
you'd have a perfect watermark

23
00:01:04.560 --> 00:01:06.880
to prove the time
when anything was recorded.

24
00:01:06.880 --> 00:01:11.760
I talked to one of the forensic experts who
works with it, and also set them a challenge. 

25
00:01:11.760 --> 00:01:16.200
- Historically, what's had to happen is:
we've had to record the grid frequency data

26
00:01:16.200 --> 00:01:20.040
ourselves by literally plugging
a recording device into the mains.

27
00:01:20.040 --> 00:01:24.600
The National Grid have recently released their data,
seven years recorded right at the source.

28
00:01:24.600 --> 00:01:27.000
It's fairly simple, actually,
signal processing.

29
00:01:27.000 --> 00:01:31.560
You're looking to see if there is a component
around 50 hertz or one of the harmonics.

30
00:01:31.560 --> 00:01:35.200
I would hesitate against advising anyone
to present themselves to court

31
00:01:35.200 --> 00:01:36.520
as a forensic expert, 

32
00:01:36.520 --> 00:01:39.800
but this type of evidence
has been used in court to date recordings.

33
00:01:39.800 --> 00:01:41.280
The longer recording you have, the better. 

34
00:01:41.280 --> 00:01:43.800
As an absolute minimum,
I think about a minute.

35
00:01:43.800 --> 00:01:47.000
If it's a noisy recording,
or if the ENF isn't particularly clear,

36
00:01:47.000 --> 00:01:49.280
then the longer recording
you have, the better. 

37
00:01:49.280 --> 00:01:52.920
- So: find the mains hum frequency in
the footage and then write code to match

38
00:01:52.920 --> 00:01:56.600
against the millions of possibilities
in the National Grid logs.

39
00:01:56.600 --> 00:01:57.400
Does it work? 

40
00:01:57.400 --> 00:01:59.000
Well to find out,
I asked some friends of mine 

41
00:01:59.000 --> 00:02:01.480
who make videos to send me over
some of their raw footage.

42
00:02:01.480 --> 00:02:03.560
Straight out of a camera, unprocessed,
unfiltered.

43
00:02:03.560 --> 00:02:07.280
And I asked the forensic team:
can you timestamp it?

44
00:02:07.280 --> 00:02:08.760
First up, footage from Taha, 

45
00:02:08.760 --> 00:02:12.440
one of the team behind educational channel
Answer in Progress.

46
00:02:12.440 --> 00:02:15.960
There's a quite strong component here, and
you can see it's got a bit of a wobble.

47
00:02:15.960 --> 00:02:18.760
We use a signal processing algorithm
to extract the peak,

48
00:02:18.760 --> 00:02:24.240
and we constrain the algorithm
so that it can't suddenly jump up here.

49
00:02:24.240 --> 00:02:28.000
So even if there is a higher amplitude,
something up here, it will

50
00:02:28.000 --> 00:02:30.840
stick to this curve. Then compare.

51
00:02:30.840 --> 00:02:31.520
If we know that 

52
00:02:31.520 --> 00:02:34.640
a recording was taken on a certain day,
then we just need to give it

53
00:02:34.640 --> 00:02:36.840
a day's worth of National Grid data.

54
00:02:36.840 --> 00:02:40.000
If we're not sure when it was taken,
we'll throw everything at it,

55
00:02:40.000 --> 00:02:43.080
and it will take a bit longer to process,
but it's not prohibitive.

56
00:02:43.080 --> 00:02:44.800
- Spot on. Got it to the second. 

57
00:02:44.800 --> 00:02:48.000
Although it's important to say
that the team never said they were certain

58
00:02:48.000 --> 00:02:51.200
about the timing, only that they had
a very high chance of being correct.

59
00:02:51.200 --> 00:02:52.680
And yes, they were. 

60
00:02:52.680 --> 00:02:56.200
Next up, footage from Hannah Witton,
sex and relationships educator.

61
00:02:56.200 --> 00:03:00.960
There wasn't a component
at around 50 hertz.

62
00:03:00.960 --> 00:03:07.160
So here we have the axis, 49 hertz to 50 hertz,
and we can see no apparent ENF trace.

63
00:03:07.160 --> 00:03:09.600
So I'm going to change the axes up
to around 100.

64
00:03:09.600 --> 00:03:12.720
And then we'll be able to see
the harmonic component around 100 hertz.

65
00:03:12.720 --> 00:03:17.640
Because it's quite strong and seems to be
behaving in a way that we'd expect ENF

66
00:03:17.640 --> 00:03:21.720
to behave, we can be fairly confident
that this is an ENF harmonic.

67
00:03:21.720 --> 00:03:23.080
What's interesting about this one is: it 

68
00:03:23.080 --> 00:03:26.840
sort of appears
halfway through the recording.

69
00:03:26.840 --> 00:03:29.720
Maybe there's a little bit of something
earlier on, but...

70
00:03:29.720 --> 00:03:32.800
So there's a possibility that maybe Hannah
was moving around when she made this recording,

71
00:03:32.800 --> 00:03:39.000
getting closer or further away from
mains powered equipment. 

72
00:03:39.000 --> 00:03:42.160
Maybe the mains hum was coming
from something on a timer or a thermostat

73
00:03:42.160 --> 00:03:45.520
in the background that just switched,
but the timestamp was accurate.

74
00:03:45.520 --> 00:03:48.160
So the next test was Steve
Mold from his science channel.

75
00:03:48.160 --> 00:03:52.600
And I thought that this one would be easy
because he sent B-roll.

76
00:03:52.600 --> 00:03:54.960
He's not talking, it's just ambient noise.

77
00:03:54.960 --> 00:03:58.800
And also: he's holding a fan
that's plugged into the mains!

78
00:03:58.800 --> 00:04:01.880
There's no strong component at all for ENF.

79
00:04:01.880 --> 00:04:04.560
So the next step would be to look at the

80
00:04:04.560 --> 00:04:08.120
other frequency regions
to see if we can identify any harmonics.

81
00:04:08.120 --> 00:04:10.720
How loud the person is often
doesn't have much of an effect.

82
00:04:10.720 --> 00:04:14.800
It's often environmental noise,
that would be lower to mask ENF.

83
00:04:14.800 --> 00:04:19.240
There is possibly something here,
but because we haven't got

84
00:04:19.240 --> 00:04:20.640
a strong trace, it's sort-of 

85
00:04:20.640 --> 00:04:24.360
fading in and out, we wouldn't
be confident to use this for ENF analysis.

86
00:04:24.360 --> 00:04:26.480
- Sometimes it just doesn't work. 

87
00:04:26.480 --> 00:04:27.240
As a last test, 

88
00:04:27.240 --> 00:04:29.160
I sent over some of my old footage
from when I was filming 

89
00:04:29.160 --> 00:04:31.200
against green screen
in my tiny flat last year. 

90
00:04:31.200 --> 00:04:33.880
I was sneaky, though,
and it hid an edit in there.

91
00:04:33.880 --> 00:04:37.760
What could happen with an edit is:
you have a sudden jump in ENF frequency.

92
00:04:37.760 --> 00:04:39.920
In that case, it's
quite likely to be an edit. 

93
00:04:39.920 --> 00:04:42.480
It's possible that the ENF trace
would just sort of stop.

94
00:04:42.480 --> 00:04:43.680
The reason for that could be 

95
00:04:43.680 --> 00:04:47.520
because it's got an edit in,
which in this case we believe it has,

96
00:04:47.520 --> 00:04:51.480
but could also be because you walked away
from the source of the ENF energy.

97
00:04:51.480 --> 00:04:55.360
- Like any forensic analysis, this isn't
some magic tool that always works.

98
00:04:55.360 --> 00:04:56.920
Compression can stop it working. 

99
00:04:56.920 --> 00:05:00.040
You need minutes of uninterrupted audio
to have enough confidence in the result.

100
00:05:00.040 --> 00:05:03.920
And sometimes there might be no background
mains hum at all.

101
00:05:03.920 --> 00:05:07.120
But I love this
because it's an accident of technology

102
00:05:07.120 --> 00:05:11.160
that's created a tool that would sound
like science fiction 20 years ago.